Analysis of Rare Categories

Cognitive Technologies Managing Editors: D. M. Gabbay J. Siekmann Editorial Board: A. Bundy J. G. Carbonell M. Pinkal H...

Author: Jingrui He

17 downloads 433 Views 1MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!

Report copyright / DMCA form

DOWNLOAD PDF

DOWNLOAD .DOCX DOWNLOAD .PPTX

Cognitive Technologies Managing Editors: D. M. Gabbay J. Siekmann Editorial Board: A. Bundy J. G. Carbonell M. Pinkal H. Uszkoreit M. Veloso W. Wahlster M. J. Wooldridge

For further volumes: http://www.springer.com/series/5216

Jingrui He

Analysis of Rare Categories

Dr. Jingrui He Machine Learning Group IBM T.J. Watson Research Center 1101 Kitchawan Road, Route 134 Yorktown Heights, NY 10598 USA Managing Editors Prof. Dov M. Gabbay Augustus De Morgan Professor of Logic Department of Computer Science King’s College London Strand, London WC2R 2LS, UK

Prof. Dr. Jörg Siekmann Forschungsbereich Deduktions- und Multiagentensysteme, DFKI Stuhlsatzenweg 3, Geb. 43 66123 Saarbrücken, Germany

Cognitive Technologies ISSN 1611-2482 ISBN 978-3-642-22812-4 e-ISBN 978-3-642-22813-1 DOI 10.1007/978-3-642-22813-1 Springer Heidelberg Dordrecht London New York Library of Congress Control Number: 2011941602 ACM Codes: I.2.6, E.1, H.2.8 © Springer-Verlag Berlin Heidelberg 2012 This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer. Violations are liable to prosecution under the German Copyright Law. The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. Printed on acid-free paper Springer is part of Springer Science+Business Media (www.springer.com)

Preface In many real world problems, rare categories (minority classes) play an essential role despite their extreme scarcity. For example, in ﬁnancial fraud detection, the vast majority of ﬁnancial transactions are legitimate, and only a small number may be fraudulent; in Medicare fraud detection, the percentage of bogus claims is small, but the total loss is signiﬁcant; in network intrusion detection, malicious network activities are hidden among huge volumes of routine network trafﬁc; in astronomy, only 0.001% of the objects in sky survey images are truly beyond the scope of current science and may lead to new discoveries; in spam image detection, near-duplicate spam images are difﬁcult to discover from the large number of non-spam images; in rare disease diagnosis, rare diseases affect less than 1 out of 2000 people, but the consequences can be very severe. Therefore, the discovery, characterization and prediction of rare categories or rare examples may protect us from fraudulent or malicious behaviors, provide aid for scientiﬁc discoveries, and even save lives. This book focuses on the analysis of rare categories, where the majority classes have a smooth distribution, and the minority classes exhibit a compactness property. Furthermore, we focus on the challenging case where the support regions of the majority and minority classes overlap each other. To the best of our knowledge, this book is the ﬁrst end-to-end investigation of rare categories. In this book, we focus on both supervised and unsupervised rare category analysis depending on the availability of the label information. In the supervised settings, our ﬁrst task is rare category detection, which is to discover at least one example from each minority class with the help of a labeling oracle. For this task, simply applying random sampling proves very expensive in terms of the number of label requests to the oracle, so we need to design more effective methods. Then, given labeled examples from all the classes, our second task is rare category characterization. The goal here is to ﬁnd a compact representation for the minority classes in order to identify all the rare examples with high precision and recall. On the other hand, in the unsupervised settings, we do not have access to a labeling v

vi

Preface oracle. Here we propose to co-select candidate examples from the minority classes and the relevant features, which beneﬁts both tasks (rare category selection and feature selection). For each of the above tasks, we have designed effective algorithms with theoretical guarantees as well as good empirical results.

Jingrui He 08/31/2011

Contents 1

2

3

Introduction 1.1 Problem Deﬁnition . 1.2 General Assumptions 1.3 Outline of This Book 1.4 General Notation . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

Survey and Overview 2.1 Active Learning . . . . . . . . . . . . . 2.2 Imbalanced Classiﬁcation . . . . . . . . 2.3 Anomaly Detection (Outlier Detection) 2.4 Clustering . . . . . . . . . . . . . . . . 2.5 Co-clustering . . . . . . . . . . . . . . 2.6 Unsupervised Feature Selection . . . . 2.7 Rare Category Detection . . . . . . . .

. . . .

. . . . . . .

. . . .

. . . . . . .

. . . .

. . . . . . .

. . . .

. . . . . . .

. . . .

. . . . . . .

. . . .

. . . . . . .

. . . .

. . . . . . .

. . . .

. . . . . . .

. . . .

. . . . . . .

. . . .

. . . . . . .

. . . .

1 2 3 4 7

. . . . . . .

9 9 11 12 13 13 15 15

Rare Category Detection 3.1 Rare Category Detection with Priors for Data with Features . 3.1.1 Rare Category Detection for the Binary Case . . . . 3.1.2 Rare Category Detection for Multiple Classes . . . . 3.1.3 Experimental Results . . . . . . . . . . . . . . . . . 3.2 Prior-free Rare Category Detection for Data with Features . 3.2.1 Semiparametric Density Estimation for Rare Category Detection . . . . . . . . . . . . . . . . . . . . 3.2.2 Algorithm . . . . . . . . . . . . . . . . . . . . . . . 3.2.3 Experimental Results . . . . . . . . . . . . . . . . . 3.3 Rare Category Detection for Graph Data . . . . . . . . . . . 3.3.1 GRADE Algorithm . . . . . . . . . . . . . . . . . . 3.3.2 GRADE-LI Algorithm . . . . . . . . . . . . . . . . 3.3.3 Experimental Results . . . . . . . . . . . . . . . . .

17 17 18 23 29 38 39 47 49 54 55 61 61 vii

viii

Contents

3.4

3.3.4 Discussion . . . . . . . . . . . . . . . . . . . . . . Summary of Rare Category Detection . . . . . . . . . . . . 3.4.1 Limitations . . . . . . . . . . . . . . . . . . . . . .

4 Rare Category Characterization 4.1 Optimization Framework . . . . . . . . . . . . . . . . 4.1.1 Pre-processing: Filtering . . . . . . . . . . . . 4.1.2 Problem Formulations . . . . . . . . . . . . . 4.2 Optimization Algorithm: RACH . . . . . . . . . . . . 4.2.1 Initialization Step . . . . . . . . . . . . . . . . 4.2.2 Projected Subgradient Method for Problem 4.3 4.2.3 RACH for Problem 4.1 . . . . . . . . . . . . . 4.3 Kernelized RACH Algorithm . . . . . . . . . . . . . . 4.4 Experimental Results . . . . . . . . . . . . . . . . . . 4.5 Summary of Rare Category Characterization . . . . . . 4.5.1 Limitations . . . . . . . . . . . . . . . . . . . 5 Unsupervised Rare Category Analysis 5.1 Optimization Framework . . . . . . . . . . . . . . 5.1.1 Additional Notation . . . . . . . . . . . . 5.1.2 Objective Function . . . . . . . . . . . . . 5.1.3 Justiﬁcation . . . . . . . . . . . . . . . . . 5.2 Partial Augmented Lagrangian Method . . . . . . . 5.3 Experimental Results . . . . . . . . . . . . . . . . 5.3.1 Synthetic Data Sets . . . . . . . . . . . . . 5.3.2 Real Data Sets . . . . . . . . . . . . . . . 5.4 Summary of Unsupervised Rare Category Analysis 5.4.1 Limitations . . . . . . . . . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . .

67 72 73

. . . . . . . . . . .

75 76 77 77 80 81 82 84 87 87 96 97

. . . . . . . . . .

99 100 100 101 102 104 107 107 112 121 123

6 Conclusion and Future Directions

125

Bibliography

129

Chapter 1

Introduction Imbalanced data sets are prevalent in real applications, i.e., some classes occupy the majority of the data set, a.k.a., the majority classes; whereas the remaining classes only have a few examples, a.k.a., the minority classes or the rare categories. For example, in ﬁnancial fraud detection, the vast majority of the ﬁnancial transactions are legitimate, and only a small number may be fraudulent [BKA+ 06]; in Medicare fraud detection, the percentage of bogus claims is small, but the total loss is signiﬁcant; in network intrusion detection, new malicious network activities are hidden among huge volumes of routine network trafﬁc [WXWC07][VW09]; in astronomy, only 0.001% of the objects in sky survey images are truly beyond the scope of current science and may lead to new discoveries [PM04]; in spam image detection, near-duplicate spam images are difﬁcult to discover from the large number of non-spam images [WJL+ 07]; in health care, rare diseases affect less than 1 out of 2000 people, but the consequences are severe. Compared with the majority classes, the minority classes are often of much greater interest to the users. The main focus of this book is the analysis of rare categories in an imbalanced data set, which is addressed from different perspectives. When dealing with highly imbalanced data sets, there are a number of challenges. For example, in ﬁnancial fraud detection, we may want to discover new types of fraud transactions from a large number of unexamined ﬁnancial transactions with the help of a domain expert. In a more abstract way, given an imbalanced data set, our goal is to discover a few examples from the minority classes when we have access to a labeling oracle. Due to the extreme scarcity of the new types of fraud transactions compared with normal transactions, simple methods such as random sampling would result in a huge number of label requests to the domain expert, which can be very expensive. Therefore, to reduce the labeling cost, we need more effective J. He, Analysis of Rare Categories, Cognitive Technologies, DOI 10.1007/978-3-642-22813-1_1, © Springer-Verlag Berlin Heidelberg 2012

1

2

1 Introduction

methods for discovering the new types of fraud transactions. Take rare disease diagnosis as another example. Given a small number of patients with a speciﬁc rare disease, how can we characterize this rare disease based on a subset of the medical measurements? With this characterization, we hope to help domain experts better understand the mechanism of this disease, distinguish it from other diseases for the best treatment, and identify potential patients with the same disease. The major challenge here is the insufﬁciency of label information, which might be alleviated by leveraging the information of health people as well as patients with similar diseases. Yet another example is in Medicare fraud detection. Due to the large number of medical claims submitted to the computer system for Medicare services, it is impossible for a domain expert to examine each of them and report suspicious ones. Therefore, an automated program that is able to detect complex fraud patterns with a high accuracy will greatly reduce the demand for human labor. The major challenge here is the lack of label information, which might be compensated by studying the properties of known fraud patterns. All the above challenges are associated with rare category analysis, where the main theme is to provide powerful tools for analyzing the rare categories in real applications.

1.1 Problem Deﬁnition In rare category analysis, we are given an imbalanced data set, which is initially unlabeled. Depending on the availability of label information, rare category analysis can be performed in a supervised or unsupervised fashion. In supervised rare category analysis, we have access to a labeling oracle, which is able to provide us with the label information of any example with a ﬁxed cost. In this case, rare category analysis can be divided into the following two tasks. T1 (Rare category detection): In this task, we start denovo, and propose initial candidates of each minority class to the labeling oracle (one candidate in each round) in an active learning fashion, hoping to ﬁnd at least one example from each minority class with the fewest total label requests. This task serves as the initial exploration step of the data set, and generates a set of labeled examples from each class, which can be used in the second task. T2 (Rare category characterization): In this task, we have labeled examples from each class (both majority and minority classes), which are obtained from the ﬁrst task, as well as a set of unlabeled examples as

1.2 General Assumptions

3

input. The goal is to ﬁnd a compact representation for the minority classes in order to identify all the rare examples with high precision and recall. This task serves as the exploitation step of the data set, and constructs a reliable classiﬁer, which can be used to identify future unseen examples from the minority classes. In unsupervised rare category analysis, we do not have such a labeling oracle, and the goal here is to address the following two problems. P1 (Rare category selection): selecting a set of examples which are likely to come from the same minority classes; P2 (Feature selection): selecting the features that are relevant to the minority classes.

1.2

General Assumptions

When dealing with examples with feature representations, we make the following general assumptions: — Smoothness assumption: the underlying distribution of each majority class is sufﬁciently smooth. — Compactness (clustering) assumption: the examples from the same minority class form a compact cluster in the feature space or feature subspace; An example of the underlying distribution where these assumptions are satisﬁed is shown in Fig. 1.1, which corresponds to a one-dimensional synthetic data set. The majority class has a Gaussian distribution with a large variance; whereas the minority classes correspond to the two lower variance peaks. The purpose of these assumptions is to make the minority classes identiﬁable. For example, if there are only one majority class and one minority class, and both have the same underlying distribution, for the task of rare category detection, no algorithms will perform better than random sampling. On the other hand, if the distribution of the majority class is very bumpy, and each bump is as narrow and sharp as the distribution of the minority class, then the minority classes cannot be discovered very efﬁciently either. Empirical studies support these assumptions. Furthermore, according to the experimental results shown in the following chapters, our algorithms developed based on these assumptions often achieve good performance for rare category analysis. Furthermore, we focus on cases where the support regions of the majority and minority classes overlap with each other in the feature space (although

4

1 Introduction

some proposed algorithms work well in both the separable and non-separable cases). Overlapping (non-separable) phenomenon can be observed in many real applications. For example, guileful fraudulent people often try to camouﬂage their transactions within normal transactions so that they can bypass current fraud detection algorithms [CPF06]; in spam image detection, ‘advanced’ spam images are deliberately made like normal images. Compared with the cases where the majority and minority classes are separable / near-separable from each other, which are targeted by most existing work in rare category analysis, such as [FM06] and [PM04], the overlapping (nonseparable) cases are more realistic and more challenging, but have not been well studied before. 0.2 0.18 0.16 0.14 0.12 0.1 0.08 0.06 0.04 0.02 0 −6

−4

−2

0

2

4

6

Figure 1.1: Underlying distribution of a one-dimensional synthetic data set: the majority class has a Gaussian distribution with a large variance; whereas the minority classes correspond to the two lower variance peaks.

1.3 Outline of This Book In this book, the analysis of rare categories is carried out in the following three parts, namely rare category detection, rare category characterization, and unsupervised rare category analysis. Rare Category Detection refers to the task of discovering examples

1.3 Outline of This Book

5

from new rare categories in an imbalanced data set with the help of a labeling oracle. Here, the following three questions need to be well addressed. 1. How to detect the rare categories in an unlabeled, imbalanced data set better than random sampling? 2. How to detect the rare categories for graph data, or relational data? 3. How to do rare category detection with the least prior information about the data set? The ﬁrst question is fundamental in rare category detection. Given an unlabeled, imbalanced data set, a naive way to ﬁnd examples from the minority classes is to randomly select examples from the data set to be labeled by the oracle until we have identiﬁed at least one example from each minority class. A major drawback of this random sampling strategy is the following. If a minority class is extremely rare, say its proportion in the data set is only 0.001%, in order to discover this minority class, the number of label requests by random sampling would be very large. Therefore, we need more effective methods to address the rare category detection challenge. The second question is similar to the ﬁrst one except that here we are interested in graph data (relational data) instead of data points with feature representations. Since graph data is very common in real applications, how to adapt rare category detection algorithms for data points with feature representations to this data type is of key importance. The third question aims at doing rare category detection with less prior information about the data set, and the ultimate goal is prior-free rare category detection, i.e., the algorithm is given no prior information about the data set, such as the number of classes, the proportions of different classes, etc. This question is more difﬁcult than the ﬁrst two, and yet quite important, since in real applications, given an unlabeled data set, it is sometimes difﬁcult to estimate the number of classes in the data set a priori, not to mention the proportions of different classes. Although both rare category detection and traditional active learning proceed by actively selecting examples to be labeled by an oracle, they are different in the following two aspects. First, in rare category detection, initially we do not have any labeled examples; whereas in traditional active learning, initially we have labeled examples from all the classes as input. Second, in rare category detection, our goal is to discover at least one example from each minority class with the fewest label requests to the oracle; whereas in traditional active learning, the goal is to improve the performance of the current classiﬁer with the fewest label requests to the oracle. Furthermore, rare category detection is a bottleneck in reducing the overall sampling complex-

6

1 Introduction

ity of active learning [BBL06, Das05]. That is, the sampling complexity of many active learning algorithms is dominated by the initial stage of ﬁnding at least one example from each class, especially the minority classes. Therefore, effective rare category detection algorithms could help reduce the sampling complexity of active learning by a large margin. The major difference between rare category detection and outlier detection is the following. In rare category detection, examples from the same minority class are often self-similar, potentially forming compact clusters in the feature space; and we assume that the support regions of the majority and minority classes are NOT separable from each other, which is more challenging than the separable cases. On the other hand, in anomaly detection (outlier detection), each anomaly (outlier) is a single data point; the anomalies (outliers) are typically scattered in the feature space; and the anomalies (outliers) are often separable from the normal points. Rare Category Characterization follows rare category detection. Given labeled examples from all the classes as input, the goal is to identify all the rare examples from the known minority classes with high precision and recall. Here, our question is: how to characterize the minority classes with a compact representation? Rare category characterization is of key importance in many applications, such as text retrieval where the number of documents relevant to a particular query is very small, and the goal is to retrieve all of them at the top of the ranked list returned to the user. Rare category characterization bears similarity but also has some differences with imbalanced classiﬁcation. In both tasks, the data set is imbalanced, and we have labeled training examples from each class. However, in imbalanced classiﬁcation, the goal is to construct a classiﬁer that optimizes a discriminative criterion for both the majority and the minority classes, such as balanced accuracy, G-mean, etc [Cha09]; whereas in rare category characterization, we only focus on the minority classes, and aim to identify all (or nearly all) the rare examples from the unlabeled data set with high precision and recall. On the other hand, in rare category characterization, we are able to obtain a compact representation for the minority classes, which can be used to improve the interpretability of the learning results to domain experts who have no knowledge in machine learning and data mining; whereas in imbalanced classiﬁcation, such representations are not provided. Unsupervised Rare Category Analysis, where we do not have access to any labeling oracle. Under certain assumptions, we can perform both rare category selection, which is to select a set of examples that are likely to come from the minority classes, and feature selection, which is to select a set of

1.4 General Notation

7

features relevant to the minority classes. Rare category selection is similar to rare category detection in that both start denovo, i.e., no label information is available at the beginning. However, in rare category detection, the labeling oracle gives the label information of the candidate example selected in each round, and the learning algorithm can adjust its model based on this information and propose new candidates, hoping to ﬁnd at least one example from each minority class with only a few label requests; whereas in rare category selection, no label information is available during the training stage, and the goal is to select a set of examples which are likely to come from the minority classes. In many imbalanced data sets, the examples from the same minority class are close to each other in some dimensions, and far apart in others. For example, in ﬁnancial fraud detection, small-amount-probing type fraud transactions are similar to each other in terms of the small amount being stolen in each transaction; however, different transactions may occur in different locations, at different times, etc. Therefore, to better describe the minority classes, we need to identify the relevant features. Notice that existing co-clustering algorithms would fail in our settings due to the following two reasons. First, we are dealing with the cases where the majority and minority classes overlap in the feature space; whereas existing co-clustering algorithms mainly work in the separable / near-separable cases. Second, in our applications, we hope to ﬁnd the features relevant to the minority classes; whereas existing co-clustering algorithms would simply ignore the rare examples due to their scarcity, and the selected features would be relevant to the majority classes only.

1.4

General Notation

For ease of future reference, here we introduce the general notation used throughout this book, which is summarized in Table 1.1. Given a set of n unlabeled examples S = {x1 , . . . , xn }, xi ∈ Rd , which come from m distinct classes, i.e., yi ∈{1, . . . , m}. Without loss of generality, assume that ni=1 xi = 0 and n1 ni=1 (xji )2 = 1, where xji is the j th feature of xi . For the sake of simplicity, assume that there is one majority class with prior p1 , which corresponds to yi = 1, and all the other classes are minority classes with priors p2 , . . . , pm , p1 pc , c = 1. Notice that in real applications, we may have multiple majority classes. If all of them satisfy the smoothness assumption introduced in Section 1.2, they can be seen as a single class for the purpose of rare category analysis. This is because if a certain example

8

1 Introduction

is from one of the majority classes, we do not care which majority class it comes from.

Symbol S n m xi xji d yi fc

Table 1.1: Notation Deﬁnition The set of unlabeled examples The number of examples in S The number of classes in S The ith unlabeled example The j th feature of xi The dimensionality of the feature space The class label of xi The probability density function of class c, c = 1, . . . , m

Let fc denote the probability density function (pdf) of class c, where c = 1, . . . , m. Based on our discussion in Section 1.2, f1 for the majority class should satisfy the smoothness assumption; whereas fc , c = 2, . . . , m for the minority classes should satisfy the compactness assumption. Other more speciﬁc notation will be introduced where needed.

Chapter 2

Survey and Overview Rare category analysis is related to many research areas, including active learning, where the goal is to improve the classiﬁcation performance with the fewest label requests to the labeling oracle; imbalanced classiﬁcation, where the goal is to construct a classiﬁer for imbalanced data sets which is able to identify the under represented classes; anomaly detection (outlier detection), which refers to the problem of ﬁnding patterns in the data that do not conform to expected behavior; clustering, which refers to the problem of grouping similar data items into clusters; co-clustering, which generally involves grouping the data from various dimensions; and unsupervised feature selection, where the goal is to select features for the sake of grouping the data without any supervision. In this chapter, we review related work in the above areas, highlighting their differences with rare category analysis. Compared with these research areas, rare category analysis is relatively new. In this chapter, we also brieﬂy introduce some existing work on rare category detection, which is the ﬁrst task in supervised rare category analysis; whereas the other tasks have not been addressed before.

2.1

Active Learning

The key idea behind active learning is that a machine learning algorithm can achieve greater accuracy with fewer training labels if it is allowed to choose the data from which it learns [Set10]. In active learning, we assume that the class labels are obtained from a labeling oracle with some cost, and under a ﬁxed budget, we hope to maximally improve the performance of the learning algorithm. According to [Set10], there are three main settings in active learning: membership query synthesis, stream-based selective sampling, and J. He, Analysis of Rare Categories, Cognitive Technologies, DOI 10.1007/978-3-642-22813-1_2, © Springer-Verlag Berlin Heidelberg 2012

9

10

2 Survey and Overview

pool-based sampling. Many early active learning algorithms belong to membership query synthesis, such as [Ang87] [Ang01] [CGJ96]. One major problem with membership query synthesis is that the synthesized queries often have no practical meanings, and thus no appropriate labels. On the other hand, with streambased active learning and pool-based sampling, the queries always correspond to real examples. Therefore, their label information can be readily provided by the oracle. In stream-based selective sampling, given an unlabeled example, the learner must decide whether to query its class label or to discard it. For example, in [CAL92], Cohn et al compute a region of uncertainty, and query examples within it; in [DE95], Dagan et al proposed committee-based sampling, which evaluates the informativeness of an example by measuring the degree of disagreement between several model variants and only queries the more informative ones. On the other hand, in pool-based sampling, queries are selected from a pool of unlabeled examples. Its major difference from stream-based selective sampling is the large amount of unlabeled data available at query time, which reveals additional information about the underlying distribution. For example, Tong et al [TKK01] proposed an active learning algorithm that minimizes the size of the version space; Mccallum and Nigam [MN98] modiﬁed the Query-by-Committee method of active learning to use unlabeled data for density estimation, and combined this with EM to ﬁnd the class labels of the unlabeled examples. It should be mentioned that in traditional active learning, initially we have labeled examples from all the classes in order to build the very ﬁrst classiﬁer, which can be improved by actively selecting the training data. On the other hand, in rare category detection, initially we do not have any labeled examples, and the goal is to discover at least one example from each minority class with the fewest label requests. Combining rare category detection and traditional active learning, it has been noticed in [BBL06] and [Das05] that if the learning algorithm starts denovo, ﬁnding the initial labeled examples from each class (i.e., rare category detection) becomes the bottleneck for reducing the sampling complexity. Furthermore, in supervised rare category analysis, following rare category detection, the second task is rare category characterization, which works in a semi-supervised fashion. In this task, in order to get a more accurate representation of the minority classes, we can make use of active learning to select the most informative examples to be added to the labeled set.

2.2 Imbalanced Classiﬁcation

2.2

11

Imbalanced Classiﬁcation

In imbalanced classiﬁcation, the goal is to construct an accurate classiﬁer that optimizes a discriminative criterion, such as balanced accuracy, G-mean, etc [Cha09]. Existing methods can be roughly categorized into three groups [Cha09], i.e., sampling-based methods [KM97][CBHK02], adapting learning algorithms by modifying objective functions or changing decision thresholds [WC03] [HYKL04], and ensemble-based methods [SKW06][CLHB03]. To be speciﬁc, in sampling-based methods, some methods under-sample the majority classes. For example, the one-sided sampling strategy proposed in [KM97] employs Tomek links [Tom76] followed by closest nearest neighbor [HAR68] to discard the majority class examples that lie in the borderline region, are noisy or redundant. In contrast, some sampling-based methods over-sample the hard examples. For example, the DataBoost-IM method proposed in [GV04] generates synthetic examples according to the hard examples identiﬁed during the boosting algorithm; the SMOTEBoost algorithm proposed in [CLHB03] applies the SMOTE algorithm [CBHK02] to create new examples from the minority class in each boosting round. Furthermore, some methods combine over-sampling the minority class and under-sampling the majority class. For example, the SMOTE algorithm combined with under-sampling [CBHK02] was proven to outperform only undersampling the majority class and varying the loss ratios; in [TZCK09], different rebalance heuristics were incorporated into SVM modeling to tackle the problem of class imbalance, including over-sampling, under-sampling, etc. Imbalanced classiﬁcation and rare category characterization bear similarity but also have some differences. On one hand, both tasks need labeled examples from all the classes as input. On the other hand, imbalanced classiﬁcation and rare category characterization have different goals as well as different output. To be speciﬁc, in rare category characterization, we only focus on the minority classes, and aim to identify all (or nearly all) the rare examples from the unlabeled data set with high precision and recall; whereas in imbalanced classiﬁcation, the goal is to construct a classiﬁer that optimizes a discriminative criterion for both the majority and minority classes. Furthermore, in rare category characterization, we are able to obtain a compact representation for the minority classes, which can be provided to domain experts for better understanding of the learning results; whereas in imbalanced classiﬁcation, such representations are not provided.

12

2 Survey and Overview

2.3 Anomaly Detection (Outlier Detection) Anomaly detection refers to the problem of ﬁnding patterns in data that do not conform to expected behavior [CBK09]. Anomalies are often referred to as outliers. According to [CBK09], the majority of anomaly detection techniques can be categorized into classiﬁcation-based, nearest neighbor-based, clustering-based, information theoretic, spectral, and statistical techniques. For example, in [BWJ01], the authors propose a method based on a technique called pseudo-Bayes estimators to enhance an anomaly detection system’s ability to detect new attacks while reducing the false alarm rate as much as possible. In [RRS00], the authors propose a novel formulation for distance-based outliers that is based on the distance of a point from its kth nearest neighbor. Then they rank each point on the basis of its distance to its k th nearest neighbor and declare the top n points in this ranking to be outliers. In [YSZ02], the authors propose the FindOut algorithm, which is an extension of the WaveCluster algorithm [SCZ98] in which the detected clusters are removed from the data and the residual instances are declared as anomalies. In [HXD05], the authors formally deﬁne the problem of outlier detection in categorical data as an optimization problem from a global viewpoint based on entropy minimization, and present a local-search heuristicbased algorithm for efﬁciently ﬁnding feasible solutions. In [DGBK07], the authors describe distributed algorithms for doing Principal Component Analysis (PCA) using random projection and sampling based techniques. Using the approximate principal components, they develop a distributed outlier detection algorithm based on the fact that the last principal component enables identiﬁcation of data points which deviate sharply from the ‘correlation structure’ of the data. And in [AY01], the authors discuss a new technique for outlier detection which is especially suited to very high dimensional data sets. The method works by ﬁnding lower dimensional projections which are locally sparse, and cannot be discovered easily by brute force techniques because of the number of combinations of possibilities. In general, anomaly detection ﬁnds individual and isolated examples that differ from a given class in an unsupervised fashion. Typically, there is no way to characterize the anomalies since they are often different from each other. There exist a few works dealing with the case where the anomalies are clustered [PKGF03]. However, they still assume that the anomalies are separable from the normal data points. On the other hand, in rare category detection, each rare category consists of a group of points, which form a compact cluster in the feature space and are self-similar. Furthermore, we are dealing with challenging cases where the support regions of the majority

2.4 Clustering

13

and minority classes overlap with each other.

2.4

Clustering

According to [Fun01], clustering refers to the grouping together of similar data items into clusters. Existing clustering algorithms can be categorized into the following two main classes [Fun01]: parametric clustering and nonparametric clustering. In general, parametric methods attempt to minimize a cost function or an optimality criterion which associates a cost to each example-cluster assignment. It can be further classiﬁed into two groups: generative models and reconstructive models. In generative models, the basic idea is that the input examples are observations from a set of unknown distributions. For example, in Gaussian mixture models [RR95], the data are viewed as coming from a mixture of Gaussian probability distributions, each representing a different cluster; in C-means fuzzy clustering [Dun73], the membership of a point is shared among various clusters. On the other hand, reconstructive methods generally attempt to minimize a cost function. For example, K-means clustering forms clusters in numeric domains, partitioning examples into disjoint clusters [DHS00]; in Deterministic Annealing EM Algorithm (DAEM) [HB97], the maximization of the likelihood function is embedded in the minimization of the thermodynamic free energy, depending on the temperature which controls the annealing process. For nonparametric methods, two good representative examples are the agglomerative and divisive algorithms, also called hierarchical algorithms [Joh67], that produce dendrograms. In unsupervised rare category analysis, one important problem we want to address is rare category selection, i.e., selecting a set of examples which are likely to come from the minority classes. General-purpose clustering algorithms do not ﬁt here because the proportions of different classes are very skewed and the support regions of the majority and minority classes overlap with each other. In this case, general-purpose clustering algorithms tend to overlook the minority classes and generate clusters within the majority classes. Therefore, we need to develop new methods for rare category selection which leverage the property of the minority classes.

2.5

Co-clustering

The idea of using compression for co-clustering can be traced back to the information-theoretic co-clustering algorithm [DMM03], where the normal-

14

2 Survey and Overview

ized non-negative contingency table is treated as a joint probability distribution between two discrete random variables that take values over the rows and columns. Then co-clustering is deﬁned as a pair of mappings from rows to row clusters and from columns to column clusters. According to information theory, the optimal co-clustering is the one that minimizes the difference between the mutual information of the original random variables and the clustered random variables. The algorithm for minimizing the above criterion intertwines both row and column clustering at all stages. Row clustering is done by assessing closeness of each row distribution, in relative entropy, to certain ‘row cluster prototypes’. Column clustering is done similarly, and this process is iterated until it converges to a local minimum. It can be theoretically proven that the proposed algorithm never increases the criterion, and gradually improves the quality of co-clustering. Although the information-theoretic co-clustering algorithm can only be applied to bipartite graphs, the idea behind this algorithm can be generalized to more than two types of heterogeneous objects. For example, in [GLZ+ 07], the authors proposed the CBGC algorithm. It aims to do collective clustering for star-shaped inter-relationships among different types of objects. First, it transforms the star-shaped structure into a set of bipartite graphs; then it formulates a constrained optimization problem, where the objective function is a weighted sum of the Rayleigh quotients on different bipartite graphs, and the constraints are that clustering results for the same type of objects should be the same. Follow-up work includes high order co-clustering [GGP07]. Another example is the spectral relational clustering algorithm proposed in [LZWY06]. Unlike the previous algorithm, this algorithm is not restricted to star-shaped structures. It is based on a general model, collective factorization on related matrices. This model simultaneously clusters multi-type interrelated objects based on both the relation and the feature information. It exploits the interactions between the hidden structures of different types of objects through the related factorizations which share matrix factors, i.e., cluster indicator matrices. The resulting spectral relational clustering algorithm iteratively updates the cluster indicator matrices using the leading eigenvectors of a specially designed matrix until convergence. More recently, the collective matrix factorization proposed by Singh and Gordon [SG08a] [SG08b] can also be used for clustering k-partite graphs. Other related work includes (1) GraphScope [SFPY07], which uses a similar information-theoretic criterion as cross-association for time-evolving graphs to segment time into homogeneous intervals; and (2) multi-way distributional clustering (MDC) [BEYM05] which was demonstrated to outper-

2.6 Unsupervised Feature Selection

15

form the previous information-theoretic clustering algorithms at the time the algorithm was proposed. At ﬁrst glance, one may apply co-clustering algorithms to simultaneously address the problem of rare category selection and feature selection in unsupervised rare category analysis. The problem here is similar to the one mentioned in Section 2.4. That is, due to the extreme skewness of the class proportions and the overlapping support regions, general-purpose coclustering algorithms may not be able to correctly identify the few rare examples or the features relevant to the rare categories; whereas we propose an algorithm for co-selecting the rare examples and the relevant features, which addresses this problem by making use of the clustering property of the minority classes.

2.6

Unsupervised Feature Selection

Generally speaking, existing feature selection methods in the unsupervised setting can be categorized as wrapper models and ﬁlter models. The wrapper models evaluate feature subsets based on the clustering results, such as the FSSEM algorithm [DB00], the mixture-based approach which extends to the unsupervised context the mutual-information based criterion [LJF02], and the ELSA algorithm [KSM00]. The ﬁlter models are independent of the clustering algorithm, such as the feature selection algorithm based on maximum information compression index [MMP02], the feature selection method using distance-based entropy [DCSL02], and the feature selection method based on Laplacian score [HCN05]. In unsupervised rare category analysis, one of the problems we want to address is feature selection, i.e., selecting a set of features relevant to the minority classes. In our settings, since the class proportions are very skewed, the general-purpose wrapper and ﬁlter methods would fail by selecting the features primarily relevant to the majority classes. Therefore, we need new feature selection methods that are tailored for rare category analysis.

2.7

Rare Category Detection

In rare category detection, the goal is to ﬁnd at least one example from each minority class with the help of a labeling oracle, minimizing the number of label requests. Researchers have developed several methods for rare category detection. For example, in [PM04], the authors assumed a mixture model to ﬁt the data, and experimented with different hint selection methods, of which

16

2 Survey and Overview

Interleaving performs the best; in [FM06], the authors studied functions with multiple output values, and used active sampling to identify an example for each of the possible output values; in [DH08], the authors presented an active learning scheme that exploits cluster structure in the data, which was proven to be effective in rare category detection; and in [VW09], the authors proposed a new approach to rare category detection based on hierarchical mean shift, where a hierarchy is created by repeatedly applying mean shift with an increasing bandwidth on the data. Different from most existing work on rare category detection, which assumes that the majority and minority classes are separable / near-separable from each other in the feature space, in Chapter 3 of this book, we target the more challenging cases where the support regions of different classes are not separable. Furthermore, besides empirical evaluations of the proposed algorithms, we also prove their effectiveness theoretically; whereas most existing algorithms do not have such guarantees.

Chapter 3

Rare Category Detection In this chapter, we focus on rare category detection, the ﬁrst task in the supervised setting. In this task, we are given an unlabeled, imbalanced data set, which is often non-separable, and have access to a labeling oracle, which is able to give us the class label of any example with a ﬁxed cost. The goal here is to discover an least one example from each minority class with the fewest label requests. The key content of this chapter can be summarized as follows. 1. Algorithms for different data types, including data with feature representations and graph data (relational data); 2. Algorithms with different amounts of prior information; 3. Theoretical analysis of the algorithms. The rest of this chapter is organized as follows. In Section 3.1, we introduce the detection algorithms with prior information for data with feature representation. The prior-free algorithm is introduced in Section 3.2. Then in Section 3.3, we present our detection algorithms for graph data, or relational data, given full prior information or partial prior information. Finally, in Section 3.4, we give a brief summary of rare category detection.

3.1

Rare Category Detection with Priors for Data with Features

In this section, we propose prior-dependent algorithms for rare category detection in the context of active learning, which are designed for data with feature representations. We typically start denovo, without category labels, though our algorithms make no such assumption. Different from existing J. He, Analysis of Rare Categories, Cognitive Technologies, DOI 10.1007/978-3-642-22813-1_3, © Springer-Verlag Berlin Heidelberg 2012

17

18

3 Rare Category Detection

methods, we aim to solve the difﬁcult cases, i.e., we do not assume separability or near-separability of the classes. Intuitively, our algorithms make use of nearest neighbors to measure local density around each example. In each iteration, the algorithms select an example with the maximum change in local density on a certain scale, and ask the oracle for its label. The algorithms stop once they have found at least one example from each class (given the knowledge of the number of classes). When the two assumptions in Section 1.2 are satisﬁed, the proposed algorithms will select examples both on the boundary and in the interior of the minority classes, and are proven to be effective theoretically. The rest of the section is organized as follows. In Subsection 3.1.1, we introduce our algorithm for the binary case and provide theoretical justiﬁcation. In Subsection 3.1.2, we discuss the more general cases where there are multiple minority classes in the data set. Finally, Subsection 3.1.3 provides some experimental results.

3.1.1

Rare Category Detection for the Binary Case

Algorithm First let us focus on the simplest case where m = 2. Therefore, p1 = 1 − p2 , and p2 1. Here, we assume that we have an estimate of the value of p2 a priori. Next, we introduce our algorithm for rare category detection based on nearest neighbors, which is presented in Alg. 1. The basic idea is to ﬁnd maximum changes in local density, which might indicate the location of a rare category. The algorithm works as follows. Given the unlabeled set S and the prior of the minority class p2 , we ﬁrst estimate the number K of minority class examples in S. Then, for each example, we record its distance from the K th nearest neighbor, which could be realized by kd-trees [Moo91]. The minimum distance over all the examples is assigned to r . Next, we draw a hyper-ball centered at each example with radius r , and count the number of examples enclosed by this hyper-ball, which is denoted ni . ni is roughly in proportion to the local density. To measure the change of local density around a certain point xi , in each iteration of Step 3, we subtract the nk of neighboring points from ni , and let the maximum value be the score of xi . The example with the maximum score is selected for labeling by the oracle. If the example is from the minority class, stop the iteration; otherwise, enlarge the neighborhood where the scores of the examples are re-calculated and continue.

3.1 Rare Category Detection with Priors for Data with Features

19

Before giving the theoretical justiﬁcation, here, we give an intuitive explanation of why the algorithm works. Assume that the minority class is concentrated in a small region and the probability density function (pdf) of the majority class is locally smooth. Firstly, since the support region of the minority class is very small, it is important to ﬁnd its scale. The r value obtained in Step 1 will be used to calculate the local density ni . Since r is based on the minimum K th nearest neighbor distance, it is never too large to smooth out changes of local density, and thus it is a good measure of the scale. Secondly, the score of a certain point, corresponding to the change in local density, is the maximum of the difference in local density between this point and all of its neighboring points. In this way, we are not only able to select points on the boundary of the minority class, but also points in the interior, given that the region is small. Finally, by gradually enlarging the neighborhood where the scores are calculated, we can further explore the interior of the support region, and increase our chance of ﬁnding a minority class example. Algorithm 1 Nearest-Neighbor-Based Rare Category Detection for the Binary Case (NNDB) Input: S, p2 1: Let K = np2 . For each example, calculate the distance to its K th nearest neighbor. Set r to be the minimum value among all the examples. 2: ∀xi ∈ S, let N N (xi , r ) = {x|x ∈ S, x − xi ≤ r }, and ni = |N N (xi , r )|. 3: for t = 1 : n do 4: ∀xi ∈ S, if xi has not been selected, then si = max (ni −nk ); otherwise, si = −∞. Query x = arg maxxi ∈S si . 6: If the label of x is 2, break. 7: end for

xk ∈N N (xi ,tr )

5:

Justiﬁcation Next we prove that if the minority class is concentrated in a small region and the pdf of the majority class is locally smooth, the proposed algorithm will repeatedly sample in the region where the rare examples occur with a high probability. First of all, we make the following speciﬁc assumptions.

20

3 Rare Category Detection Assumptions 1. f2 (x) is uniform within a hyper-ball B of radius r centered at b, i.e., f2 (x) = V 1(r) , if x ∈ B; and 0 otherwise, where V (r) ∝ rd is the volume of B.

κ1 p 2 , ∀x ∈ 2. f1 (x) is bounded and positive in B 1 , i.e., f1 (x) ≥ (1−p 2 )V (r) κ2 p2 d B and f1 (x) ≤ (1−p2 )V (r) , ∀x ∈ R , where κ1 , κ2 > 0 are two constants. With the above assumptions, we have the following lemma and theorem. Note that variants of the following proof apply if we assume a different minority class distribution, such as a tight Gaussian. Lemma 1. ∀, δ > 0, if n ≥ max{ 2κ12 p2 log 3δ , 2(1−21−d )2 p2 log 3δ , 4 V (1r2 )4

log 3δ }, where r2 =

r

1 (1+κ2 ) d

1 2

2

2

, and V ( r22 ) is the volume of a hyper-ball with

radius r22 , then with probability at least 1 − δ, r22 ≤ r ≤ r and | nni − E( nni )| ≤ V (r ), 1 ≤ i ≤ n, where V (r ) is the volume of a hyper-ball with radius r . Proof. First, notice that the expected proportion of points falling inside B, E( |N Nn(b,r)| ) ≥ (κ1 + 1)p2 , and that the maximum expected proportion of points falling inside any hyper-ball of radius 2−d p2 . Then

r2 2,

max[E(

x∈Rd

|N N (x, n

r2 )| 2

)] ≤

ni r2 ni or ∃xi ∈ S s.t., | − E( )| > V (r )] 2 n n r2 ≤ Pr[r > r] + Pr[r < ] 2 ni r2 ni and ∃xi ∈ S s.t., | − E( )| > V (r )] + Pr[r ≥ 2 n n r2 ≤ Pr[|N N (b, r)| < K] + Pr[max |N N (x, )| > K] d 2 x∈R r2 ni ni + n Pr[| − E( )| > V (r )|r ≥ ] n n 2 N N (x, r22 ) N N (b, r) | < p2 ] + Pr[max | | > p2 ] = Pr[| n n x∈Rd ni r2 ni + n Pr[| − E( )| > V (r )|r ≥ ] n n 2 2 2 −d 2 2 2 2 ≤ e−2nκ1 p2 + e−2n(1−2 ) p2 + 2ne−2n V (r )

Pr[r > r or r <

Notice that here we are only dealing with the hard case where f1 (x) is positive within B. In the separable case where the support regions of the two classes do not overlap, we can use other methods to detect the minority class, such as the one proposed in [PM04]. 1

21

3.1 Rare Category Detection with Priors for Data with Features

where the last inequality is based on the Hoeffding bound. 2 2 −d 2 2 2 By setting e−2nκ1 p2 ≤ 3δ , e−2n(1−2 ) p2 ≤ 3δ and 2ne−2n V (r ) ≤ r2 2 2 2ne−2n V ( 2 ) ≤ 3δ , we obtain n ≥ 2κ12 p2 log 3δ , n ≥ 2(1−21−d )2 p2 log 3δ , and n ≥

4 V

1 r ( 22 )4

1 2

log 3δ .

2

Based on Lemma 1, we get the following theorem, which shows the effectiveness of the proposed algorithm. Theorem 1. If 1. Let B 2 be the hyper-ball centered at b with radius 2r. The minimum distance between the points inside B and the ones outside B 2 is not too large, i.e., min{xi − xk |xi , xk ∈ S, xi − b ≤ r, xk − b > 2r} ≤ α, where α is a positive parameter. 2. f1 (x) is locally smooth, i.e., ∀x, y ∈ Rd , |f1 (x) − f1 (y)| ≤ βx−y , α p2 OV (

r2

,r)

r2 2 where β ≤ 22d+1 V (r) 2 and OV ( 2 , r) is the volume of the overlapping region of two hyper-balls: one is of radius r, the other one is of radius r2 2 , and its center is on the sphere of the bigger one. 3. The number of examples is sufﬁciently large, i.e., n ≥ max{ 2κ12 p2 log 3δ , 2(1−21−d )2 p2 log 3δ , (1−p )4 β14 V ( r2 )4 log 3δ }. 1 2

2

2α r2

2

2

then with probability at least 1 − δ, after iterations, NNDB will query at least one example whose probability of coming from the minority class is 2d − at least 13 , and it will continue querying such examples until the ( p2 (1−p 2) α th 2) · r iteration. Proof. Based on Lemma 1, using condition 3, if the number of examples is sufﬁciently large, then with probability at least 1 − δ, r22 ≤ r ≤ r and | nni − E( nni )| ≤ (1 − p2 )βV (r ), 1 ≤ i ≤ n. According to condition 2, ∀xi , xk ∈ S s.t., xi − b > 2r, xk − b > 2r and xi − xk ≤ α, E( nni ) and E( nnk ) will not be affected by the minority class, and |E( nni )−E( nnk )| ≤ (1−p2 )βV (r ) ≤ (1−p2 )βV (r). Note that α is always bigger than r. Based on the above inequalities, we have |

ni nk ni ni nk nk ni nk − | ≤ | − E( )| + | − E( )| + |E( ) − E( )| n n n n n n n n ≤ 3(1 − p2 )βV (r) (3.1)

From Inequality 3.1, it is not hard to see that ∀xi , xk ∈ S, s.t., xi − b > 2r and xi − xk ≤ α, nni − nnk ≤ 3(1 − p2 )βV (r), i.e., when tr = α, si ≤ 3(1 − p2 )βV (r) n

(3.2)

22

3 Rare Category Detection

This is because if xk − b ≤ 2r, the minority class may also contribute to nk n , and thus the score may be even smaller. On the other hand, based on condition 1, there exist two points xk , xl ∈ S, s.t., xk − b ≤ r, xl − b > 2r, and xk − xl ≤ α. Since the conr

p2 ·OV ( 22 ,r) , so E( nnk ) − V (r) r r p ·OV ( 2 ,r) p ·OV ( 2 ,r) − (1 − p2 )βV (r ) ≥ 2 V (r)2 − (1 − p2 )βV (r). E( nnl ) ≥ 2 V (r)2 ni ni Since for any example xi ∈ S, we have | n − E( n )| ≤ (1 − p2 )βV (r ) ≤

tribution of the minority class to E( nnk ) is at least

(1 − p2 )βV (r), therefore

p2 · OV ( r22 , r) nk nl − ≥ − 3(1 − p2 )βV (r) n n V (r) p2 · OV ( r22 , r) 3(1 − p2 )p22 · OV ( r22 , r) ≥ − V (r) 2d+1 V (r) Since p2 is very small, p2 p2 )βV (r), i.e., when tr = α,

3(1−p2 )p22 ; 2d+1

therefore,

sk > 3(1 − p2 )βV (r) n

nk n

−

nl n

> 3(1 −

(3.3)

In Step 4 of the proposed algorithm, we gradually enlarge the neighborhood to calculate the change of local density. When tr = α, based on inequalities (2) and (3), ∀xi ∈ S, xi − b > 2r, we have sk > si . Therefore, in this round of iteration, we will pick an example from B 2 . In order for tr to be equal to α, the value of t would be rα ≤ 2α r2 . If we further increase t so that tr = wα, where w > 1, we have the following conclusion: ∀xi , xk ∈ S, s.t., xi − b > 2r and xi − xk ≤ wα, nk ni si n − n ≤ (w + 2)(1 − p2 )βV (r), i.e., n ≤ (w + 2)(1 − p2 )βV (r). (w+2)(1−p )p2

2 2 2 As long as p2 ≥ , i.e., w ≤ p2 (1−p − 2, then ∀xi ∈ S, 2d 2) xi − b > 2r, sk > si , and we will pick examples from B 2 . Since r ≤ r, 2d − the algorithm will continue querying examples in B 2 until the ( p2 (1−p 2) α th 2) · r iteration. Finally, we show that the probability of picking a minority class example from B 2 is at least 13 . To this end, we need to calculate the maximum probability mass of the majority class within B 2 . Consider the case where the maximum value of f1 (x) occurs at b, and this pdf decreases by β every time x moves away from b in the direction of the radius by α, i.e., the shape of f1 (x) is a cone in (d+1) dimensional space. Since f1 (x) must integrate to 1, 1 (b) i.e., V ( αfβ1 (b) )· fd+1 = 1, where V ( αfβ1 (b) ) is the volume of a hyper-ball with d

3.1 Rare Category Detection with Priors for Data with Features 1

23

d+1 β d+1 . Therefore, the probability radius αfβ1 (b) , we have f1 (b) = ( Vd+1 (α) ) 2 mass of the majority class within B is: d

2r β 2r β) + V (2r) < V (2r)f1 (b) α α d+1 1 d 1 d V (r) d + 1 d+1 d+1 β d+1 ) β d+1 = 2d = V (2r)( 1 (d + 1) V (α) V (α) d+1

V (2r)(f1 (b) −

1

d

1

< (d + 1) d+1 (2d+1 V (r)β) d+1 ≤ (d + 1) d+1 (

p22 · OV ( r22 , r) d ) d+1 < 2p2 V (r)

where V (2r) is the volume of a hyper-ball with radius 2r. Therefore, if we select a point at random from B 2 , the probability that this point is from the p2 p2 ≥ p2 +2p = 13 . minority class is at least p2 +(1−p 2 2 )·2p2

3.1.2

Rare Category Detection for Multiple Classes

In many real applications, there are often multiple minority classes2 . Therefore, we need to develop an algorithm that is able to discover examples from all the minority classes. Algorithm In Subsection 3.1.1, we have discussed rare category detection for the binary case. In this subsection, we focus on the case where m > 2. To be speciﬁc, let p1 , . . . , pm be the priors of the m classes, and p1 pc , c = 1. Our goal is to use as few label requests as possible to ﬁnd at least one example from each class. The NNDB algorithm proposed in Subsection 3.1.1 can be easily generalized to multiple classes, which is presented in Alg. 2. It works as follows. Given the priors for the minority classes, we ﬁrst estimate the number Kc of examples from class c in the set S. Then, for class c, at each example, we record its distance from the Kcth nearest neighbor. The minimum distance over all the examples is the class speciﬁc radius, and is assigned to rc . Next, we draw a hyper-ball centered at example xi with radius rc , and count the number of examples enclosed by this hyper-ball, which is denoted nci . nci is roughly in proportion to the local density. To ﬁnd examples from class c, in each iteration of Step 10, we subtract the local density of neighboring points from nci , and let the maximum value be the score of xi . The example with 2 As discussed in Section 1.4, multiple majority classes can be seen as a single majority class if they all satisfy the smoothness assumption introduced in Section 1.2

24

3 Rare Category Detection

the maximum score is selected for labeling by the oracle. If the example is from class c, stop the iteration; otherwise, enlarge the neighborhood where the scores of the examples are re-calculated and continue. Justiﬁcation As before, next we prove that if the minority classes are concentrated in small regions and the pdf of the majority class is locally smooth, the proposed algorithm will repeatedly sample in the regions where the rare examples occur with a high probability. First of all, we make the following speciﬁc assumptions. Assumptions 1. The pdf fc (x) of minority class c is uniform within a hyper-ball Bc of 1 , if x ∈ Bc ; radius rc 3 centered at bc , c = 2, . . . , m, i.e., fc (x) = V (r c) and 0 otherwise, where V (rc ) ∝ rcd is the volume of Bc .

2. f1 (x) is bounded and positive in Bc , c = 2, . . . , m, i.e., f1 (x) ≥ κc1 pc κc2 pc d p1 V (rc ) , ∀x ∈ Bc and f1 (x) ≤ p1 V (rc ) , ∀x ∈ R , where κc1 , κc2 > 0 are two constants.4 rc Furthermore, for each minority class c, c = 2, . . . , m, let rc2 = 1 ; (1+κc2 ) d

and let OV ( r2c2 , rc ) be the volume of the overlapping region of two hyperballs: one is of radius rc ; the other one is of radius r2c2 , and its center is on the sphere of the previous one. To prove the performance of the proposed ALICE algorithm, we ﬁrst have the following lemma. Lemma 2. For each minority class c, c = 2, . . . , m, ∀c , δc > 0, if n ≥ 3m−3 1 max{maxm c=2 2κ2 p2 log δ ,

maxm c=2

c1 c

1 2(1−2−d )2 p2c

ability at least 1 − δ,

1 log 3m−3 r δ }, then with prob4 V ( c2 )4 2 nci nci rc and | n − E( n )| ≤ V (rc ), 1 ≤ j ≤ n.

m log 3m−3 δ , maxc=2 rc2 2

≤ rc ≤

Proof. First, notice that for each minority class c, the expected proportion of points falling inside Bc , E( |N N (bnc ,rc )| ) ≥ (κc1 + 1)pc , and that the maximum expected proportion of points falling inside any hyper-ball of radius This is the actual radius, as opposed to the class speciﬁc radius rc . Notice that here we are only dealing with the hard case where f1 (x) is positive within Bc . In the separable case where the support regions of the majority and the minority classes do not overlap, we can use other methods to detect the minority classes, such as the one proposed in [PM04]. 3 4

3.1 Rare Category Detection with Priors for Data with Features

25

Algorithm 2 Active Learning for Initial Class Exploration (ALICE) Input: S, p2 , . . . , pm 1: Initialize all the minority classes as undiscovered. 2: for c = 2 : m do 3: Let Kc = npc , where n is the number of examples. 4: For each example, calculate the distance between this example and its Kcth nearest neighbor. Set rc to be the minimum value among all the examples. 5: end for 6: for c = 2 : m do 7: ∀xi ∈ S, let N N (xi , rc ) = {x|x ∈ S, x − xi ≤ rc }, and nci = |N N (xi , rc )|. 8: end for 9: for c = 2 : m do 10: If class c has been discovered, continue. 11: for t = 2 : n do 12: For each xi that has been selected, sci = −∞; for all the other max (nci − nck ). examples, sci = xk ∈N N (xi ,trc )

Select and query the label of x = arg maxxi ∈S sci . 14: If the label of x is equal to c, break; otherwise, mark the class that x belongs to as discovered. 15: end for 16: end for 13:

rc2 2 ,

max[E(

x∈Rd

|N N (x, n

rc2 )| 2

)] ≤ 2−d pc . Then

Pr[∃c, s.t., rc > rc or ∃c, s.t., rc <

rc2 or ∃c, ∃xi ∈ S 2

nc nci − E( i )| > V (rc )] n n m m rc2 Pr[rc > rc ] + Pr[rc < ] ≤ 2

s.t., |

c=2

+ ≤

m

c=2 m

c=2

rc2 nc nc Pr[rc ≥ and ∃xi s.t., | i − E( i )| > V (rc )] 2 n n Pr[|N N (bc , rc )| < Kc ] +

c=2

+

m c=2

m c=2

n Pr[|

nci n

− E(

nci n

Pr[max |N N (x, x∈Rd

)| > V (rc )|rc ≥

rc2 ] 2

rc2 )| > Kc ] 2

26

3 Rare Category Detection m

m N N (x, r2c2 ) N N (bc , rc ) = Pr[| Pr[max | | < pc ] + | > pc ] n n x∈Rd c=2 c=2

+n

m

Pr[|

c=2

≤

m

e

nci rc2 nc − E( i )| > V (rc )|rc ≥ ] n n 2

−2nκ2c1 p2c

c=2

+

m

e

−2n(1−2−d )2 p2c

+ 2n

c=2

m

e−2n

2V

(rc )2

c=2

where the last inequality is based on the Hoeffding bound. 2 2 −d 2 2 2 δ δ , e−2n(1−2 ) pc ≤ 3m−3 and 2ne−2n V (rc ) By setting e−2nκc1 pc ≤ 3m−3 2V

≤ 2ne−2n

(

rc2 2 ) 2

log 3m−3 δ , and n ≥

≤

δ 3m−3 , we obtain n 1 log 3m−3 r δ . 4 V ( c2 )4

≥

1 2κ2c1 p2c

log 3m−3 δ ,n ≥

1 2(1−2−d )2 p2c

2

Based on Lemma 2, we get the following theorem, which shows the effectiveness of the proposed algorithm. Theorem 2. If 1. For minority class c, c = 2, . . . , m, let Bc2 be the hyper-ball centered at bc with radius 2rc . The minimum distance between the points inside Bc and the ones outside Bc2 is not too large, i.e., maxm c=2 min{xi − xk |xi , xk ∈ S, xi − bc ≤ rc , xi − bc > 2rc } ≤ α. 2. The minority classes are far apart, i.e., if xi , xk ∈ S, xi − bc ≤ rc , xk − bc ≤ rc , c, c = 2, . . . , m, and c = c , then xi − xk > α. 3. f1 (x) is locally smooth, i.e., ∀x, y ∈ Rd , |f1 (x) − f1 (y)| ≤ βx−y , α where β ≤ minm c=2

,rc ) p2c OV ( c2 2 . 2d+1 V (rc )2 r

4. The number of examples is sufﬁciently large, i.e., n ≥ max{maxm c=2 1 1 3m−3 3m−3 1 m m log , max log , max r 2 −d 2 2 4 2 c2 c=2 2(1−2 ) p c=2 p β 4 V ( δ δ 2κ p )4 c1 c

3m−3 δ }.

c

1

2

log then with probability at least 1 − δ, in every iteration of Step 9, after r2α c2 rounds of Step 11, ALICE will query at least one example whose probability of coming from a minority class is at least 13 . Proof. Based on Lemma 2, using condition 4, let = p1 β. If the number of examples is sufﬁciently large, then with probability at least 1 − δ, for each nc nc minority class c, c = 2, . . . , m, r2c2 ≤ rc ≤ r and | ni −E( ni )| ≤ p1 βV (rc ), 1 ≤ i ≤ n. To prove the theorem, given a point xi ∈ S, we say that xi is ‘far from all the minority classes’ iff for every minority class c, xi − bc > 2rc , i.e., xi is not within Bc2 . According to condition 3, ∀xi , xk ∈ S s.t., xi and xk

27

3.1 Rare Category Detection with Priors for Data with Features

nc

nc

are far from all the minority classes and xi − xk ≤ α, E( ni ) and E( nk ) will not be affected by the minority classes. Therefore, in iteration i of Step nc nc 9 where we aim to ﬁnd examples from minority class c, |E( ni ) − E( nk )| ≤ p1 βV (rc ) ≤ p1 βV (rc ). Furthermore, since α is always bigger than rc , nc nc nc nc nc nc nc nc | i − k | ≤ | i −E( i )|+| k −E( k )|+|E( i )−E( i )| ≤ 3p1 βV (rc ) n n n n n n n n (3.4) From Inequality 3.4, it is not hard to see that ∀xi , xk ∈ S, s.t., xi is far from nc nc all the minority classes and xi − xk ≤ α, ni − nk ≤ 3p1 βV (rc ), i.e., when trc = α, sci (3.5) ≤ 3p1 βV (rc ) n This is because if xk is not far from any of the minority classes, the minority nc classes may also contribute to nk , so the score of xi may be even smaller. On the other hand, based on conditions 1 and 2, there exist two points xu , xv ∈ S, s.t., xu − bc ≤ rc , xv is far from all the minority classes, c and xu − xv ≤ α. Since the contribution of minority class c to E( nnu ) r

r

c c pc ·OV ( c2 ,rc ) pc ·OV ( c2 ,rc ) 2 2 , so E( xnu ) − E( xnv ) ≥ − p1 βV (rc ) V (rc ) V (rc ) rc2 pc ·OV ( 2 ,rc ) nc − p1 βV (rc ). Since for any example xi ∈ S, we have | ni V (rc ) nc E( ni )| ≤ p1 βV (rc ) ≤ p1 βV (rc ), therefore pc · OV ( r2c2 , rc ) nu nv − ≥ − 3p1 βV (rc ) n n V (rc ) pc · OV ( r2c2 , rc ) 3p1 p2c · OV ( r2c2 , rc ) ≥ − V (rc ) 2d+1 V (rc )

is at least

2

1 pc Since pc is very small, pc 6p ; therefore, nnu − nnv > 2d+1 3p1 βV (rc ), i.e., when trc = α, scu > 3p1 βV (rc ) n c

c

3p1 p2c ·OV ( c2 ,rc ) 2 2d+1 V (rc ) r

≥ −

≥

(3.6)

In Step 11 of the proposed method, we gradually enlarge the neighborhood to calculate the change of local density to continue seeking an example of the minority class. When trc = α, based on Inequalities 3.5 and 3.6, ∀xi ∈ S s.t., xi is far from all the minority classes, we have scu > sci . Therefore, in this round of iteration, we will pick an example that is NOT far from one of the minority classes, i.e., there exists a minority class ct s.t., the selected example is within Bc2t . Note that ct is not necessarily equal to c, which is the minority class that we would like to discover in Step 9 of the method. Finally, we show that the probability of picking an example that belongs to minority class ct from Bc2t is at least 13 . To this end, we need to calculate

28

3 Rare Category Detection

the maximum probability mass of the majority class within Bc2t . Consider the case where the maximum value of f1 (x) occurs at bct , and this pdf decreases by β every time x moves away from bct in the direction of the radius by α, i.e., the shape of f1 (x) is a cone in (d + 1) dimensional space. Since f1 (x) αf (b ) f (b ) αf (b ) must integrate to 1, i.e., V ( 1β ct ) · 1d+1ct , where V ( 1β ct ) is the vol1

αf (b )

d

d+1 β d+1 . ume of a hyper-ball with radius 1β ct , we have f1 (bct ) = ( Vd+1 (α) ) Therefore, the probability mass of the majority class within Bc2t is: 2rct β 2rct β) + V (2rct ) V (2rct )(f1 (bct ) − α α d+1 1 d d + 1 d+1 ) β d+1 < V (2rct )f1 (bct ) = V (2rct )( V (α) 1 d 1 d V (rct ) d+1 β d+1 < (d + 1) d+1 (2d+1 V (r )β) d+1 = 2d ct 1 (d + 1) (V (α)) d+1 r 1 p2 · OV ( c2t 2 , rct ) d ) d+1 < 2pct ≤ (d + 1) d+1 ( ct V (rct )

where V (2rct ) is the volume of a hyper-ball with radius 2rct . Therefore, if we select a point at random from Bc2t , the probability that this point is from p pct = 13 . minority class ct is at least pc +pc1t·2pc ≥ pc +2p c t

t

t

t

Implementation Issues According to our theorem, in each iteration of Step 9, with high probability, we may pick examples belonging to the rare classes after selecting a small number of examples. However, the discovered rare class ct may not be the same as the rare class c that we hoped to discover in this iteration of Step 9. Furthermore, we may repeatedly select examples from class ct before ﬁnding one example from class c. To address these issues, we have modiﬁed the original ALICE algorithm to produce MALICE, which is shown in Alg. 3. There are two major differences between MALICE and ALICE. (1) In Step 13 of MALICE, once we have labeled an example, any unlabeled example within the class speciﬁc radius of this example will be precluded from selection. Since we have proved that with high probability, the class speciﬁc radius is less than the actual radius, this modiﬁcation will help prevent examples of the same class from being selected repeatedly. (2) In Step 16 of MALICE, if the labeled example belongs to a rare class other than class c, we will not enlarge the neighborhood based on which the scores of the examples are re-calculated. This is to increase the chance that if trc is close to α, we will select examples from Bc2 .

3.1 Rare Category Detection with Priors for Data with Features

29

Algorithm 3 Modiﬁed Active Learning for Initial Class Exploration (MALICE) Input: S, p2 , . . . , pm 1: Initialize all the rare classes as undiscovered. 2: for c = 2 : m do 3: Let Kc = npc . 4: For each example, calculate the distance between this example and its Kcth nearest neighbor. Set rc to be the minimum value among all the examples. 5: end for 6: Let r1 = maxm c=2 rc . 7: for c = 2 : m do 8: ∀xi ∈ S, let N N (xi , rc ) = {x|x ∈ S, x − xi ≤ rc }, and nci = |N N (xi , rc )|. 9: end for 10: for c = 2 : m do 11: If class c has been discovered, continue. 12: for t = 2 : n do 13: For each xi that has been selected, ∀xk ∈ S, s.t., xi − xk ≤ ry i , sck = −∞; for all the other examples, sci = max (nci − nck ). xk ∈N N (xi ,trc )

14: 15: 16: 17: 18:

Select and query the label of x = arg maxxi ∈S sci . If the label of x is equal to c, break; otherwise, mark the class that x belongs to as discovered. If x is from a new class other than c, t = t − 1. end for end for

3.1.3

Experimental Results

In this subsection, we compare our algorithms (NNDB and MALICE) with the best method proposed in [PM04] (Interleave) and random sampling (RS) on both synthetic and real data sets. In Interleave, we use the number of classes as the number of components in the mixture model. For both Interleave and RS, we run the experiments 100 times and report the average. Synthetic Data Sets Fig. 3.1(a) shows a synthetic data set where the pdf of the majority class is Gaussian and the pdf of the minority class is uniform within a small hyper-

30

3 Rare Category Detection

ball. There are 1000 examples from the majority class and only 10 examples from the minority class. Using Interleave, we need to label 35 examples, using RS, we need to label 101 examples, and using NNDB, we only need to label 3 examples in order to sample one from the minority class, which are denoted as ‘x’s in Fig. 3.1(b). Notice that the ﬁrst 2 examples that NNDB selects are not from the correct region. This is because the number of examples from the minority class is very small, and the local density may be affected by the randomness in the data. In Fig. 3.2(a), the X-shaped data consisting of 3000 examples correspond to the majority class, and the four characters ‘NIPS’ correspond to four minority classes, which consist of 138, 79, 118, and 206 examples respectively. Using Interleave, we need to label 1190 examples, using RS, we need to label 83 examples, and using MALICE, we only need to label 5 examples in order to get one from each of the minority classes, which are denoted as ‘x’s in Fig. 3.2(b). Notice that in this example, Interleave is even worse than RS. This might be because some minority classes are located in the region where the density of the majority class is not negligible, and thus may be ‘explained’ by the majority-class mixture-model component.

Real Data Sets In this subsection, we compare different methods on two real data sets: Abalone [AN07] and Shuttle [AN07]. The ﬁrst data set consists of 4177 examples, described by 7-dimensional features. The examples come from 20 classes: the proportion of the largest class is 16.50%, and the proportion of the smallest class is 0.34%. For the second data set, we sub-sample the original training set to produce a smaller data set with 4515 examples, described by 9-dimensional features. The examples come from 7 classes: the proportion of the largest class is 75.53%, and the proportion of the smallest class is 0.13%. The comparison results are shown in Fig. 3.3 and Fig. 3.4 respectively. Further, Table 3.1 and Table 3.2 show the conﬁdence intervals for the number of label requests needed to discover a certain number of classes in the data set for different methods. This is obtained via bootstrapping5 . 5

Notice that different runs of bootstrapping are not independent of each other. However, since MALICE is a deterministic method, bootstrapping may be the only way to obtain its conﬁdence intervals given that we only have one sample from each domain.

3.1 Rare Category Detection with Priors for Data with Features 5

4

3

2

1

0

−1 −3

−2

−1

0

1

2

3

4

(a) Data set: there are 1000 examples from the majority class, denoted as blue dots, and only 10 examples from the minority class, denoted as red balls. 5

4

3

2

1

0

−1 −3

−2

−1

0

1

2

3

(b) Examples selected by NNDB, denoted as green ‘x’s.

Figure 3.1: Synthetic data set 1.

4

31

32

3 Rare Category Detection

200 180 160 140 120 100 80 60 40 20 0 0

50

100

150

200

250

(a) Data set: there are 3000 examples from the majority class, denoted as blue dots; there are 138, 79, 118, and 206 examples from each minority class, denoted as red balls.

200 180 160 140 120 100 80 60 40 20 0 0

50

100

150

200

(b) Examples selected by MALICE, denoted as green ‘x’s.

Figure 3.2: Synthetic data set 2.

250

3.1 Rare Category Detection with Priors for Data with Features

33

Table 3.1: Conﬁdence intervals for the number of label requests needed by different methods in the Abalone data set. No. of Classes MALICE Interleave RS No. of Classes MALICE Interleave RS No. of Classes MALICE Interleave RS No. of Classes MALICE Interleave RS No. of Classes MALICE Interleave RS

2 2±0.00 2±0.00 2±0.00 6 13±2.20 38±11.37 11±4.18 10 25±5.80 53±25.95 22±12.79 14 57±13.03 89±49.69 67±35.47 18 108±29.02 206±119.71 145±101.53

3 3±0.00 4±0.50 4±0.75 7 14±3.10 39±14.90 12±4.55 11 31±8.12 76±29.50 23±16.74 15 59±13.46 95±61.92 71±42.81 19 114±35.74 248±152.71 273±138.06

4 4±0.00 7±2.96 5±0.45 8 15±3.62 40±16.76 14±6.94 12 55±10.10 78±34.75 40±21.87 16 61±16.50 108±80.80 104±54.43

5 6±0.42 37±9.11 8±2.25 9 18±4.58 42±21.29 15±9.11 13 56±11.84 88±39.11 63±28.50 17 62±17.06 138±94.93 136±73.05

20 125±40.03 270±196.23 483±165.23

Table 3.2: Conﬁdence intervals for the number of label requests needed by different methods in the Shuttle data set. No. of Classes 2 3 4 5 MALICE 6±2.61 17±3.95 48±8.64 49±9.92 Interleave 2±0.00 52±9.90 107±16.15 109±30.59 RS 7±2.85 9±5.14 13±8.82 63±27.51 No. of Classes 6 7 MALICE 76±5.77 87±12.64 Interleave 115±45.30 140±58.83 RS 100±57.92 512±109.56

34

3 Rare Category Detection

Table 3.3: Conﬁdence intervals for the number of label requests needed by MALICE with different priors in the Abalone data set. No. of Classes MALICE −20% −10% −5% +20% +10% +5% No. of Classes MALICE −20% −10% −5% +20% +10% +5% No. of Classes MALICE −20% −10% −5% +20% +10% +5% No. of Classes MALICE −20% −10% −5% +20% +10% +5% No. of Classes MALICE −20% −10% −5% +20% +10% +5%

2 2±0.00 2±0.00 2±0.00 2±0.00 2±0.00 2±0.00 2±0.00 6 13±2.20 14±3.05 16±2.62 13±2.11 10±2.16 12±2.14 14±2.23 10 25±5.80 26±7.36 29±6.32 26±6.63 32±6.62 17±5.87 21±6.38 14 57±13.03 66±16.37 59±14.23 58±13.11 41±11.40 51±12.39 53±13.74 18 108±29.02 74±34.68 114±30.04 107±29.29 80±28.08 94±34.57 95±31.84

3 3±0.00 3±0.00 3±0.00 3±0.00 3±0.00 3±0.00 3±0.00 7 14±3.10 15±3.69 17±3.20 14±3.16 11±3.18 13±3.24 15±3.22 11 31±8.12 31±9.22 35±8.06 31±8.65 38±7.56 20±7.96 26 ±7.58 15 59±13.46 70±19.54 60±14.56 60±17.08 60±13.80 53±15.64 55±16.27 19 114±35.74 135±52.07 117±43.54 109±40.29 82±35.04 101±36.38 97±39.92

4 4±0.00 5±0.25 5±0.19 5±0.96 5±0.02 4±0.00 4±0.00 8 15±3.62 16±3.84 22±4.16 15±4.00 15±4.05 14±3.78 16±4.01 12 55±10.10 64±11.46 57±10.71 56±10.70 39±8.07 49±10.38 51±9.03 16 61±16.50 72±22.21 62±19.30 62±18.90 73±16.32 88±16.96 56±16.77

5 6±0.42 8±2.18 13±1.86 7±1.47 6±0.62 9±1.55 13±1.60 9 18±4.58 24±4.62 23±5.30 19±4.72 20±4.65 16±4.78 17±5.29 13 56±11.84 65±13.96 58±13.18 57±11.33 40±8.87 50±12.47 52±10.59 17 62±17.06 73±21.80 65±17.77 63±17.53 77±23.88 93±25.83 57±21.88

20 125±40.03 230±61.22 126±46.66 116±51.17 141±41.34 171±48.17 105±48.60

35

3.1 Rare Category Detection with Priors for Data with Features

Notice that the number of classes in the two tables always starts from 2, since with all the methods, the ﬁrst label request always results in a new class being discovered. From these results, we have the following observations. First, when the number of classes discovered is small, the average number of label requests by RS is smaller than MALICE; however, we are only interested in the classes whose proportions are much smaller than the rest, and to discover those classes, RS needs many more label requests compared with MALICE. Second, the performance of Interleave is not as good as MALICE in terms of the average number of label requests on the two data sets. This is because as the number of components becomes larger, the mixture model generated by Interleave is less reliable due to the lack of labeled examples, thus we need to select more examples. Furthermore, the majority and minority classes may not be near-separable, which is a disaster for Interleave. On the other hand, MALICE does not assume a generative model for the data, and only focuses on the change in local density, which is more effective on the two data sets. Finally, the conﬁdence intervals given by MALICE are always signiﬁcantly smaller than those by both Interleave and RS, showing that the performance of MALICE is much more stable. 20 18

Classes Discovered

16 14 12

MALICE

10

Interleave RS

8 6 4 2 0 0

100

200

300

400

Number of Selected Examples Figure 3.3: Learning curve for Abalone data set.

500

36

3 Rare Category Detection

Imprecise Priors The proposed algorithms need the priors of the minority classes as input. In this subsection, we test the robustness of MALICE against modest misestimations of the class priors. The performance of NNDB is similar to MALICE so we omit the results here. In the experiments, we use the same data sets as in Subsubsection 3.1.3, and add/subtract 5%, 10%, and 20% from the true priors of the minority classes. The results are shown in Fig. 3.5 and Fig. 3.6. Further, Table 3.3 and Table 3.4 show the conﬁdence intervals for the number of label requests needed by MALICE with different priors to discover a certain number of classes in the data set. This is obtained via bootstrapping. From these results, we can see that MALICE is very robust to small perturbations in the priors. For example, on the Abalone data set, if we subtract 10% from the true priors, only one more label request is needed in order to ﬁnd all the classes.6 Furthermore, small perturbations in the priors do not affect the width of the conﬁdence intervals by a large margin. 7

Classes Discovered

6

5 MALICE 4

Interleave RS

3

2

1 0

100

200

300

400

500

600

Number of Selected Examples Figure 3.4: Learning curve for Shuttle data set. 6

Interestingly, the performance of MALICE is sometimes improved with inaccurate priors. This might be attributed to accidental selection of examples from new classes in Step 15 of MALICE, which are not the target class c.

3.1 Rare Category Detection with Priors for Data with Features

37

20 18

Classes Discovered

16 14 12 −5% −10% −20% 0 +5% +10% +20%

10 8 6 4 2 0 0

50

100

150

200

250

Number of Selected Examples

Figure 3.5: Robustness study on Abalone data set: -5%, -10%, and -20% denote the performance of MALICE after we subtract 5%, 10%, and 20% from the true priors; +5%, +10%, and +20% denote the performance of MALICE after we add 5%, 10%, and 20% to the true priors.

7

Classes Discovered

6

5

4 −5% −10% −20% 0 +5% +10% +20%

3

2

1 0

20

40

60

80

100

Number of Selected Examples

Figure 3.6: Robustness study on Shuttle data set: -5%, -10%, and -20% denote the performance of MALICE after we subtract 5%, 10%, and 20% from the true priors; +5%, +10%, and +20% denote the performance of MALICE after we add 5%, 10%, and 20% to the true priors.

38

3 Rare Category Detection

Table 3.4: Conﬁdence intervals for the number of label requests needed by MALICE with different priors to discover a certain number of classes in the Shuttle data set. No. of Classes 2 3 4 5 MALICE 6±2.61 17±3.95 48±8.64 49±9.92 −20% 7±3.08 10±4.36 47±9.26 48±11.37 −10% 7±2.76 18±4.07 52±8.28 53±10.48 −5% 7±2.71 18±4.10 50±8.10 51±9.98 +20% 5±2.45 9±3.36 28±7.30 46±7.52 +10% 7±2.18 8±3.43 46±7.80 47±8.61 +5% 7±2.31 8±3.77 47±8.12 48±9.26 No. of Classes 6 7 MALICE 76±5.77 87±12.64 −20% 74±6.60 85±14.40 −10% 80±6.70 91±14.62 −5% 78±6.19 89±13.87 +20% 65±9.90 76±9.85 +10% 71±7.51 82±12.38 +5% 72±7.72 83±12.07

3.2 Prior-free Rare Category Detection for Data with Features In the last section, we have discussed rare category detection algorithms that need prior information about the data set as input, including the number of classes and the proportions of different classes. However, in some real applications, it is often the case that we do not know the number of classes in the data set, not to mention the proportions of different classes. For example, in outbreaks of epidemics, the number of classes (indicating the number of different diseases) and the class proportions could vary drastically over time as new diseases occur, strain or die out. To address this problem, in this section, we focus on the more challenging case where we do not have any prior information about the data set. The proposed method, SEmiparametric Density Estimation-based Rare category detection (SEDER), implicitly performs semiparametric density estimation using specially designed exponential families, and selects the examples whose gradients have the largest norm for labeling by the oracle. In this way, it focuses on the areas with

3.2 Prior-free Rare Category Detection for Data with Features

39

the maximum change in the local density. Different from existing methods, SEDER does not require any prior information about the data set. Therefore, it is more suitable for real applications. The rest of the section is organized as follows. In Subsection 3.2.1, we introduce the specially designed exponential families used in SEDER and derive the scoring function. The complete algorithm of SEDER is presented in Subsection 3.2.2. Finally, in Subsection 3.2.3, we compare SEDER with state-of-the-art techniques on both synthetic and real data sets.

3.2.1

Semiparametric Density Estimation for Rare Category Detection

In rare category detection, based on our assumptions introduced in Section 1.2, abrupt changes in local density indicate the presence of rare classes. By sampling in these areas, we have a high probability of ﬁnding examples from the rare classes. Following this line of reasoning, our proposed method SEDER implicitly estimates the density using specially designed exponential families, which essentially deﬁne a semiparametric model. At each data point, we set the score to be the norm of the gradient of the estimated density, which measures the maximum change rate of the local density, and pick the examples with the largest scores to be labeled by the oracle. Although the intuition of SEDER and ALICE (MALICE) is quite similar: to pick the examples with the maximum change in the local density, ALICE (MALICE) is a nearest-neighbor-based method, it depends on the proportions of different classes to set the size of the neighborhood, and the scores of the examples roughly indicate the change in the local density; whereas SEDER is based on semiparametric density estimation, it is prior-free, i.e., it does not require any prior information about the data set, and the scores measure exactly the maximum change rate in the local density. Additional Notation Besides the general notation introduced in Section 1.4, we further deﬁne the following additional notation in Table 3.5. Specially Designed Exponential Families Traditional density estimation methods belong to two categories [ET96]: by ﬁtting a parametric model via maximum likelihood, or by nonparametric methods such as kernel density estimation. For the purpose of rare category detection, parametric models are not appropriate since we can not assume

40

3 Rare Category Detection

a speciﬁc form of the underlying distribution for a given data set. On the other hand, the estimated density based on nonparametric methods tends to be under-smoothed, and examples from rare classes will be buried among numerous spikes in the estimated density.

Symbol gβ (x) g0 (x) β0 t(x) tj (x) β1 β1j σj β βˆ l(β) j gβ (xj ) g j (xj ) bj ˆbj A B C Di (x) si

Table 3.5: Notation Deﬁnition The density deﬁned by specially designed exponential families The carrier density The normalizing parameter in gβ (x) The p × 1 vector of sufﬁcient statistics The j th component of t(x) The p × 1 parameter vector The j th component of β1 The bandwidth for the j th feature (β1 , β0 ) The maximum likelihood estimate of β The log-likelihood of the data The marginal distribution of the j th feature based on gβ (x) The true marginal distribution of the j th feature Positive parameter which is a function of β1j The maximum likelihood estimate of bj 1 n

n

k=1

(σ j )2 1 n n 1 n

j j (x −x )2 i k )(xji )2 j 2(σ )2 j j 2 n (x −x ) i k i=1 exp(− 2(σ j )2 )

n

i=1

exp(−

j 2 k=1 (xk ) d √ 1 j=1 2πbj σ j

exp(−

(xj −bj xji )2 ) 2(σ j )2 bj

The score of xi

As proposed in [ET96], these two kinds of methods can be combined by putting an exponential family through a kernel density estimator, the socalled specially designed exponential families. It is a favorable compromise between parametric and nonparametric density estimation: the nonparametric smoother allows local adaptation to the data, while the exponen-

3.2 Prior-free Rare Category Detection for Data with Features

41

tial term matches some of the data’s global properties, and makes the density much smoother [ET96]. To be speciﬁc, the estimated density gβ (x) = g0 (x) exp(β0 + β1T t(x)) [ET96]. Here, x ∈ Rd , g0 (x) is a carrier density, t(x) is a p × 1 vector of sufﬁcient statistics, β1 is a p × 1 parameter vector, and β0 is a normalizing parameter that makes gβ (x) integrate to 1. In our application, we use the kernel density estimator with the Gaussian kernel j (xj −x )2 1 exp(− 2(σ j )i2 ), as the carrier density, i.e., g0 (x) = n1 ni=1 dj=1 √2πσ j where xj is the j th feature of x, xji is the j th feature of the ith data point, and σ j is the bandwidth for the j th feature. In SEDER, σ j is determined by cross validation [Sco92] on the j th feature. Here, the parameters β = (β1 , β0 ) can be estimated according to the following theorem. Theorem 3. The maximum likelihood estimate βˆ of β satisﬁes the following conditions [ET96]: ∀j ∈ {1, . . . , p}

x1

···

tj (x)gβˆ(x)dxd · · · dx1 =

xd

1 j t (xi ) n n

i=1

where tj (x) is the j th component of the vector t(x). Proof. Firstly, notice that β0 is a normalizing parameter that makes gβ (x) integrate to 1, i.e., β0 = − log

x1

···

xd

g0 (x) exp(β1T t(x))dxd · · · dx1

Therefore, ∀j ∈ {1, . . . , p} ∂β0 ∂β1j

=− =− =−

j T d 1 x1· · · xdt (x)g0 (x) exp(β1 t(x))dx · · · dx T d 1 x1 · · · xd g0 (x) exp(β1 t(x))dx · · · dx x1

x1

··· ···

xd xd

tj (x)g0 (x) exp(β0 + β1T t(x))dxd · · · dx1 tj (x)gβ (x)dxd · · · dx1

where β1j is the j th component of the vector β1 . n Secondly, the log-likelihood of the data is l(β) = i=1 log(gβ (xi )) = n n T i=1 log(g0 (xi )) + nβ0 + i=1 β1 t(xi ). Taking the partial derivative of l(β) with respect to β1j , we have:

42

3 Rare Category Detection

∂l(β)

=n

∂β1j

∂β0 ∂β1j

+

x1

tj (xi )

i=1

= −n

n

···

xd

tj (x)gβ (x)dxd · · · dx1 +

n

tj (xi )

i=1

Setting the partial derivative to 0, we have that the maximum likelihood estimate βˆ of β satisﬁes x1 · · · xd tj (x)gβˆ(x)dxd · · · dx1 = n1 ni=1 tj (xi ), ∀j ∈ {1, . . . , p}. In SEDER, we set the vector of sufﬁcient statistics to be t(x) = [(x1 )2 , . . . , (xd )2 ]T 7 . If we estimate the parameters according to Theorem 3, different parameters will be coupled due to the normalizing parameter β0 . Let β1j be the j th component of the vector β1 . In order to de-couple the estimation of different β1j s, we make the following changes. Firstly, we decompose β0 into β0j s such that dj=1 β0j = β0 , then gβ (x) can be seen as a kernel density estimator with a ‘special’ kernel, i.e.,

(xj − xji )2 1 1 [√ exp(− ) exp(β0j + β1j (xj )2 )] j )2 j n 2(σ 2πσ n

gβ (x) =

d

i=1 j=1

Next, we relax the constraint on β0j s, and let them depend on xji in such a way that √ xj

(xj − xji )2 1 j exp(− ) exp(β0i + β1j (xj )2 )dxj = 1 j )2 j 2(σ 2πσ

(3.7)

j where β0i implies the dependence of β0j on xji . In this way, the marginal distribution of the j th feature is

gβj (xj ) 7

=

x1

···

xj−1

xj+1

···

xd

gβ (x)dxd · · · dxj+1 dxj−1 · · · dx1

Note that the following analysis also applies to other forms of the sufﬁcient statistics, such as t(x) = [x1 , . . . , xd ]T . In all our experiments, the second order sufﬁcient statistics perform the best. So we use this form in SEDER.

43

3.2 Prior-free Rare Category Detection for Data with Features (xj − xji )2 1 1 j = {√ exp(− ) exp(β0i + β1j (xj )2 )· j 2 j n 2(σ ) 2πσ i=1 1 (xk − xki )2 k √ exp(− ) exp(β0i + β1k (xk )2 )dxk } k )2 k 2(σ k 2πσ k=j x n

(xj − xji )2 1 1 j √ exp(− ) exp(β0i + β1j (xj )2 ) n 2(σ j )2 2πσ j n

=

i=1

To estimate the parameters in our current model, we have the following theorem. j j of β1j and β0i Theorem 4. The maximum likelihood estimates βˆ1j and βˆ0i satisfy the following conditions: ∀j ∈ {1, . . . , d} n

(xjk )2 =

k=1

n

n

k=1

where Eij ((xj )2 ) =

xj (x

(xj −xj )2

j j 2 i k ˆj i=1 exp(β0i − 2(σ j )2 )Ei ((x ) ) n (xjk −xji )2 ˆj i=1 exp(β0i − 2(σ j )2 )

j )2 √ 1 2πσ j

exp(−

(xj −xji )2 j ) exp(βˆ0i + βˆ1j (xj )2 )dxj . 2(σ j )2

j Proof. First of all, according to Equation 3.7, we have β0i = − log

exp(− j ∂β0i

∂β1j

(xj −xji )2 ) exp(β1j (xj )2 )dxj . 2(σ)2

=−

xj

=− =

√ 1 2πσ j

exp(−

(3.8)

xj

√ 1 2πσ j

Therefore,

(xj −xji )2 ) exp(β1j (xj )2 )(xj )2 dxj 2(σ)2 j

(xj −xi )2 ) exp(β1j (xj )2 )dxj 2(σ)2 (xj − xji )2 1 j √ exp(− ) exp(β0i + β1j (xj )2 )(xj )2 dxj 2(σ j )2 2πσ j

xj

√ 1 2πσ j

exp(−

xj −Eij ((xj )2 )

Then the log-likelihood of the data on the j th component is l(β1j )

=

n k=1

=

n k=1

n

log(gβj (xjk )) (xjk − xji )2 1 1 j √ log( exp(− ) · exp(β0i + β1j (xjk )2 )) j 2 j n 2(σ ) 2πσ n

i=1

(xj − xj )2 1 j log( √ exp(β1j (xjk )2 ) · exp(− k j 2i ) exp(β0i )) = j 2(σ ) n 2πσ i=1 k=1 n

44

3 Rare Category Detection Taking the partial derivative of l(β1j ) with respect to β1j , we have: ∂l(β1j ) ∂β1j

=

n

(xjk )2 +

k=1

=

n

n

n

j i=1 exp(β0i −

n

k=1

(xjk )2

−

k=1

n

n

j i=1 exp(β0i

j i=1 exp(β0i

n

−

j (xjk −xji )2 ∂β0i ) j 2 2(σ ) ∂β j 1

−

(xjk −xji )2 ) 2(σ j )2

(xjk −xji )2 )Eij ((xj )2 ) 2(σj )2

j i=1 exp(β0i −

k=1

(xjk −xji )2 ) 2(σ j )2

Setting the partial derivative to 0, we have that the maximum likeli j j hood estimates βˆ1j and βˆ0i of β1j and β0i respectively satisfy nk=1 (xjk )2 = n

n

j j (x −x )2 i k )Eij ((xj )2 ) j 2(σ )2 j 2 j n ˆj (xk −xi ) i=1 exp(β0i − 2(σ j )2 )

i=1

k=1

j exp(βˆ0i −

.

Notice that according to Theorem 4, the β1j s can be estimated separately, which greatly simpliﬁes our problem. At the ﬁrst glance, Equation 3.8 is hard to solve. Next, we let β1j = (1 − b1j ) 2(σ1j )2 , where bj = 1 is a positive parameter, the introduction of which will simplify this equation. According j can be expressed in terms of bj , i.e., to Equation 3.7, β0i

j β0i

(xj − xji )2 1 exp(− + β1j (xj )2 )dxj 2(σ j )2 2πσj xj (xj )2 + bj (xji )2 − 2bj xj xji 1 √ = − log exp(− )dxj 2(σ j )2 bj 2πσj xj (1 − bj )(xji )2 1 = − log bj 2(σ j )2 2 = − log

√

Therefore, the estimated density becomes (xj − bj xji )2 1 1 √ exp(− ) n 2(σ j )2 bj 2πbj σ j i=1 j=1 n

g˜b (x) =

d

(3.9)

j Replacing βˆ1j and βˆ0i with functions of ˆbj (the maximum likelihood estimate of bj ) in the deﬁnition of Eij ((xj )2 ), we have Eij ((xj )2 ) = ˆbj (σ j )2 + (ˆbj )2 (xji )2 , and Equation 3.8 becomes n n (xjk )2 = nˆbj (σj )2 + (ˆbj )2 k=1

k=1

n

(xjk −xji )2 (1−ˆbj )(xji )2 − )(xji )2 j 2 2(σ ) 2(σj )2 n (xjk −xji )2 (1−ˆbj )(xji )2 − 2(σ j )2 ) i=1 exp( 2(σ j )2

i=1 exp(

45

3.2 Prior-free Rare Category Detection for Data with Features

In general, the value of βˆ1j is very close to 0, and gβˆ(x) is a smoothed version of g0 (x). Therefore, ˆbj should be close to 1, and we can re-write the above equation as follows. 1 1 j 2 ˆj j 2 (xk ) ≈ b (σ ) + (ˆbj )2 n n n

n

k=1

k=1

(xjk −xji )2 j 2 i exp(− 2(σj )2 )(xi )

(xjk −xji )2 i exp(− 2(σ j )2 )

This is a second-degree polynomial equation of ˆbj , and the roots can be easily obtained by Vieta’s theorem8 , i.e., ∀j ∈ {1, . . . , d} √ 2 ˆbj = −B + B + 4AC (3.10) 2A where A =

1 n

n

j j (x −x )2 i k )(xji )2 2(σ j )2 j j n (x −x )2 i k i=1 exp(− 2(σ j )2 )

n

i=1

k=1

(xjk )2 .

exp(−

, B = (σ j )2 , and C =

Theorem 5. Let g j (xj ) be the true density for the j th feature. If dg j (xji ) dxji

1 n

≥ −1 + O(1), then ˆbj ≤ 1 and βˆ1j ≤ 0.

1 n

n

k=1

n

xji i=1 g j (xj ) · i

Proof. For the sake of simplicity, let z = xj , h = σ j , and f (z) = gj (xj ). Then 1 A= n n

B=h

k=1 2

n

(zk −zi )2 2 i=1 exp(− 2h2 )(zi ) n 2 (zk −zi ) i=1 exp(− 2h2 )

1 (zk )2 n n

C=

k=1

Consider the following regression problem where the true regression function r(z) = z 2 , the noise has mean 0, and we use kernel regression to estimate this function. Then A − C is the bias of kernel regression on the )f (zi ) 2 ) z k(z)dz + training data, i.e., A − C = n1 ni=1 h2 ( 12 r (zi ) + r (zfi(z i) 2 O(h ) [Was05], where k(z) is the Gaussian kernel used in kernel regression, 8

Note that the other root

−B−

√

B 2 +4AC 2A

is disregarded since it is negative.

46

3 Rare Category Detection

2 2 n 2zi f (zi ) exp(− z2 ). Therefore, A − C = h2 + hn i=1 f (zi ) + O(h2 ), and A + B − C ≥ 0 if and only if n1 ni=1 ziff(z(zi )i ) ≥ −1 + O(1). √ B 2 +4AC ≤ Given that A + B − C ≥ 0, we can show that ˆbj = −B+ 2A √ 2 −B+ B +4A(A+B) = 1, and βˆ1j = (1 − ˆ1j ) 2(σ1j )2 ≤ 0. 2A

i.e., k(z) =

√1 2π

b

In Section 1.2, we have made the following assumptions: 1) the distribution of the majority classes is sufﬁciently smooth; and 2) the minority classes form compact clusters in the feature space. In this case, the ﬁrst order derivative of the density would be close to 0 for most examples, and have large absolute values for a few examples near the rare classes. Therefore, the condition in Theorem 5 is always satisﬁed, and the exponential term appended to the carrier density decreases away from the origin. Scoring Function Once we have estimated all the parameters using Equation 3.10, we can measure the change in the local density at each data point based on the estimated density in Equation 3.9. Note that at each data point, if we pick a different direction, the change in local density is different. In SEDER, we measure the change along the gradient, which gives the maximum change at each data point. Theorem 6. Using the estimated density in Equation 3.9, ∀x ∈ Rd , the d (ni=1 Di (x)(xl −bl xli ))2 maximum change rate of the density at x is , l=1 ((σ l )2 bl )2 (xj −bj xj )2 where Di (x) = n1 dj=1 √ 1 j j exp(− 2(σj )2 bij ) is the contribution of xi 2πb σ to the estimated density at x. Proof. For x ∈ Rd , let the gradient vector be w ∈ Rd . We have ∀l ∈ {1, . . . , d} (xj −bj xj )2

i n d ∂˜ gb (x) 1 xl − bl xli exp(− 2(σj )2 bj ) √ = (− ) wl = ∂xl n (σ l )2 bl 2πbj σ j i=1 j=1

=−

n Di (x)(xl − bl xl ) i=1

(σ l )2 bl

where wl is the lth component of w.

i

3.2 Prior-free Rare Category Detection for Data with Features

47

Therefore, the maximum change rate of the density at x is

d n

Di (x)(xl − bl xli ) 2 ) w2 = (− (σ l )2 bl i=1 l=1

d n l l l 2

( i=1 Di (x)(x − b xi )) = ((σ l )2 bl )2 l=1

If the distribution of the majority class is sufﬁciently smooth, and the minority classes form compact clusters in the feature space, the minority classes are always located in the regions where the density changes the most. Therefore, in SEDER, to discover the rare classes, we set the score of each example to be the maximum change rate of the density at this example, i.e., ∀k ∈ {1, . . . , n}

d n l l l 2

( i=1 Di (xk )(xk − b xi )) (3.11) sk = ((σ l )2 bl )2 l=1

where sk is the score of xk . We pick the examples with the largest scores for labeling until we ﬁnd at least one example from each class.

3.2.2

Algorithm

The intuition of SEDER is to select the examples with the maximum change in the local density for labeling by the oracle. As introduced in Subsubsection 3.2.1, the scores of the examples measure the maximum change rate in the local density, and they do not take into account the fact that nearby examples tend to have the same class label. Therefore, if we ask the oracle to label all the examples with large scores, we may repeatedly select examples from the most distinctive rare class, rather than discovering all the rare classes. To address this problem in SEDER, we make use of the following heuristic: if xi ∈ S has been labeled, ∀xk ∈ S, xk = xi , if ∀j ∈ {1, . . . , d}, |xji − xjk | ≤ 3σ j , we preclude xk from being selected. In other words, if an unlabeled example is very close to a previously labeled one, it is quite likely that the labels of the two examples are the same, and labeling that example will not give a high probability of detecting a new rare class. The size of the neighborhood is set to 3σ j such that the estimated density for the examples outside this neighborhood using the Gaussian kernel is hardly affected

48

3 Rare Category Detection

by the labeled example. It should be pointed out that the feedback strategy is orthogonal to the remaining parts of the proposed algorithm. In our experiments, we ﬁnd that despite its simplicity, the current strategy leads to satisfactory performance. Algorithm 4 SEmiparametric Density Estimation based Rare category detection (SEDER) Input: Unlabeled data set S Output: The set I of selected examples and the set L of their labels 1: Initialize I = φ and L = φ. 2: for j = 1 : d do 3: Calculate the bandwidth σ j using cross validation [Sco92]. 4: Calculate the maximum likelihood estimate ˆbj of the parameter bj according to Equation 3.10. 5: end for 6: for i = 1 : n do 7: Calculate the score si of the ith example according to Equation 3.11 using the estimated parameters. 8: end for 9: while the labeling budget is not exhausted do 10: Set S = {x|x ∈ S, ∀i ∈ I, ∃j ∈ {1, . . . , d}, s.t., |xj − xji | > 3σ j } 11: Query x = argmaxxi ∈S si for its label yx 12: I = I ∪ {x}, L = L ∪ {yx }; 13: end while The proposed method, SEDER, is summarized in Alg. 4. It works as follows. Firstly, we initialize the set I of selected examples and the set L of their labels to empty sets. Then step 2 to step 5 calculate the parameters in our model. Step 6 to step 8 calculate the score for each example in S. Finally, step 9 to step 13 gradually include the example with the maximum score into I and its label into L until we run out of the labeling budget. In each round, the selected example should be far away from all the labeled examples. Note that: 1) unlike the methods proposed in Section 3.1, SEDER does not need to be given the number of classes in S or any other information, hence it is more suitable for real applications; 2) in SEDER, we do not need to explicitly calculate the density at each example; 3) SEDER does not depend on the assumption that the majority and minority classes are separable or near-separable.

3.2 Prior-free Rare Category Detection for Data with Features

3.2.3

49

Experimental Results

In this subsection, we compare SEDER with MALICE proposed in Section 3.1, Interleave (the best method proposed in [PM04]), random sampling (RS) and SEDER with bj = 1 for j = 1, . . . , d (abbreviated as Kernel, which is equivalent to using the kernel density estimator to estimate the density and to get the scores) on both synthetic and real data sets. For this purpose, we run these methods until all the classes have been discovered, and compare the number of label requests by each method in order to ﬁnd a certain number of classes. Note that SEDER, MALICE and Kernel are deterministic, whereas the results for Interleave and random sampling are averaged over 100 runs. Here we would like to emphasize that only SEDER, RS and Kernel do not need any prior information about the data set, whereas MALICE and Interleave need extra information about the data set as inputs, such as the number of classes and the proportions of different classes. When such prior information is not available, which is quite common in real applications, MALICE and Interleave are not applicable. Synthetic Data Sets Fig. 3.7(a) shows the underlying distribution of a 1-dimensional synthetic data set. The majority class with 2000 examples has a Gaussian distribution with a large variance; whereas the minority classes with 50 examples each correspond to the two lower-variance peaks. As can be seen from this ﬁgure, the ﬁrst two examples selected by SEDER (green ‘x’s) are both from the regions where the density changes the most. Fig. 3.7(b) shows a 2-dimensional synthetic data set. The majority class has 2000 examples (blue dots) with a Gaussian distribution. The four minority classes (red balls) all have different shapes, and each has 267, 280, 84 and 150 examples respectively. This data set is similar to the one used in [HC08]. To discover all the classes, SEDER only needs to label 6 examples, which are represented by green ‘x’s in the ﬁgure; whereas random sampling needs to label more than 50 examples on average. Real Data Sets In this subsubsection, we present the experimental results on some real data sets. The properties of the data sets are summarized in Table 3.6. Notice that all these data sets are skewed: the proportion of the smallest class is less than 5%. For the last three data sets (Page Blocks, Abalone and Shuttle), it is even less than 1%. We refer to these three data sets as ‘highly’ skewed; and the remaining two (Ecoli and Glass) are referred to as ‘moderately’ skewed.

50

3 Rare Category Detection 0.2 0.18 0.16 0.14 0.12 0.1 0.08 0.06 0.04 0.02 0 −6

−4

−2

0

2

4

6

(a) Underlying distribution of a one-dimensional synthetic data set: the majority class has a Gaussian distribution with a large variance; whereas the minority classes correspond to the two lower variance peaks.

200 180 160 140 120 100 80 60 40 20 0 −50

0

50

100

150

200

250

(b) Two-dimensional synthetic data set: there are 2000 examples from the majority class, denoted as blue dots; there are 267, 280, 84, and 150 examples from each minority class, denoted as red balls.

Figure 3.7: Synthetic data sets: examples selected by SEDER are denoted as green ‘x’s.

3.2 Prior-free Rare Category Detection for Data with Features

51

First, let us focus on the ‘moderately’ skewed data sets, which are shown in Fig. 3.8 and Fig. 3.9. On the Ecoli data set, to discover all the classes, MALICE needs 36 label requests, Interleave needs 41 label requests on average, RS needs 43 label requests on average, Kernel needs 78 label requests, and SEDER only needs 20 label requests; on the Glass data set, to discover all the classes, MALICE needs 18 label requests, Interleave needs 24 label requests on average, RS needs 31 label requests on average, Kernel needs 102 label requests, and SEDER needs 22 label requests. Therefore, on these ‘moderately’ skewed data sets, the performance of SEDER is better than or comparable with MALICE, which requires more prior information than SEDER including the number of classes in the data set and the proportions of different classes. Next, let us look at the ‘highly’ skewed data sets. For example, in the Shuttle data set, the largest class has 580 times more examples than the smallest class. On the Page Blocks data set (Fig. 3.10), to discover all the classes, SEDER needs 36 label requests, MALICE needs 23 label requests, Interleave needs 77 label requests on average, RS needs 199 label requests on average, and Kernel needs more than 1000 label requests; on the Abalone data set (Fig. 3.11), to discover all the classes, SEDER needs 316 label requests, MALICE needs 179 label requests, Interleave needs 333 label requests on average, RS needs 483 label requests on average9 , and Kernel needs more than 1000 label requests; on the Shuttle data set (Fig. 3.12), to discover all the classes, SEDER needs 249 label requests, MALICE needs 87 label requests, Interleave needs 140 label requests on average, RS needs 512 label requests on average, and Kernel needs more than 1000 label requests. Based on the above results, we have the following observations. First, SEDER, RS and Kernel require no prior information about the data set, and yet SEDER is signiﬁcantly better than RS and Kernel in all the experiments. Second, the performance of Interleave is sometimes worse than SEDER (except Fig. 3.12), even though it is given additional information about the number of classes in the data set. Finally, although MALICE is better than SEDER for these ‘highly’ skewed data sets, in real applications, it is very difﬁcult to estimate the number of classes in the data set, not to mention the proportions of the different classes. If the information provided to MALICE is not accurate enough, the performance of MALICE may be negatively affected. Moreover, when such information is not available, MALICE is not applicable at all. 9 Note that on the Abalone data set, the results of MALICE and Interleave are slightly different from [HC07]. This is due to the effect of normalization on the data.

52

3 Rare Category Detection

Proportion of Classes Discovered

1 0.9 0.8 0.7 0.6 0.5

SEDER MALICE

0.4

RS Interleave

0.3

Kernel 0.2 0.1 0

10

20

30

40

50

60

70

80

Number of Selected Examples

Figure 3.8: Ecoli data set.

Proportion of Classes Discovered

1 0.9 0.8 0.7 0.6 SEDER 0.5

MALICE RS

0.4

Interleave Kernel

0.3 0.2 0.1 0

20

40

60

80

Number of Selected Examples

Figure 3.9: Glass data set.

100

120

53

3.2 Prior-free Rare Category Detection for Data with Features

Proportion of Classes Discovered

1 0.9 0.8 0.7 SEDER

0.6

MALICE RS

0.5

Interleave Kernel

0.4 0.3 0.2 0

20

40

60

80

100

120

140

160

180

200

Number of Selected Examples

Figure 3.10: Page Blocks data set.

1

Proportion of Classes Discovered

0.9 0.8 0.7 0.6 0.5

SEDER MALICE

0.4

RS Interleave

0.3

Kernel 0.2 0.1 0 0

500

1000

Number of Selected Examples

Figure 3.11: Abalone data set.

1500

54

3 Rare Category Detection

Proportion of Classes Discovered

1 0.9 0.8 0.7 0.6 SEDER

0.5

MALICE RS

0.4

Interleave 0.3

Kernel

0.2 0.1 0

50

100

150

200

250

300

350

400

450

500

550

Number of Selected Examples

Figure 3.12: Shuttle data set.

Table 3.6: Properties of the data sets used Data Set Ecoli [AN07] Glass [AN07] Page Blocks [AN07] Abalone [AN07] Shuttle [AN07]

n 336 214 5473 4177 4515

d 7 9 10 7 9

m 6 6 5 20 7

Largest Class 42.56% 35.51% 89.77% 16.50% 75.53%

Smallest Class 2.68% 4.21% 0.51% 0.34% 0.13%

3.3 Rare Category Detection for Graph Data In the previous two sections, we have discussed rare category detection algorithms for data with feature representations. In many real applications, sometimes the data is given to us as a graph, such as transaction networks and social networks, and we hope to discover minority classes on the graph data, such as collusion-type fraud transactions and groups of terrorists who constantly interact with each other. To address this problem, in this section, we propose a new rare category detection method for graph data, or rela-

3.3 Rare Category Detection for Graph Data

55

tional data: Graph-based Rare Category Detection (GRADE). The basic idea is to utilize the global similarity matrix and get more compact clusters for the examples from the minority classes, which is motivated by the manifold ranking algorithm [ZWG+ 03] and the consistency method [ZBL+ 03]. This results in sharp changes in local density near the boundary of the minority classes and thus makes it easier to discover those classes. Furthermore, we improve the GRADE algorithm to get the GRADE-LI algorithm, which requires less prior information compared with the GRADE algorithm, and thus is more suitable for real applications. Notice that our algorithms can deal with both data with feature representations and graph data, whereas existing rare category detection methods can only work with data with feature representations. The rest of this section is organized as follows. In Subsection 3.3.1, we describe the GRADE algorithm for rare category detection and analyze its effectiveness. The improved algorithm GRADE-LI is presented in Subsection 3.3.2. In Subsection 3.3.3, we show some experimental results of GRADE and GRADE-LI compared with existing methods. Finally, we discuss some implementation issues in Subsection 3.3.4.

3.3.1

GRADE Algorithm

In this subsection, we ﬁrst introduce the GRADE algorithm for data with feature representation. Then we discuss its application to graph data. Algorithm The GRADE algorithm for examples with observed features is described in Alg. 5. Here α is a positive parameter provided to the algorithm which is very close to 1. The algorithm works as follows. First of all, we calculate the maximum number of examples K in each minority class. Then using this number, we pick the parameter σ, which is the smallest distance to the K th nearest neighbor. Next, we construct the pair-wise similarity matrix W , whose elements are calculated using the Gaussian kernel. In Step 4, we construct the diagonal matrix D, whose elements are the row sums of W . Next, we calculate the normalized matrix W and the global similarity matrix A. The speciﬁc form of the global similarity matrix has been used in the manifold ranking algorithm [ZWG+ 03] for ranking data with respect to a query point and the consistency method [ZBL+ 03] for semi-supervised learning. The following steps are based on the similarity measure in A. For each class c, we calculate

56

3 Rare Category Detection

the number of examples Kc from this class, and ﬁnd the largest global similarity to the (Kc )th nearest neighbor, which is the class speciﬁc similarity ac . Then, for each example xi , we ﬁnd all of its neighbors with global similarity bigger than or equal to ac , which is denoted N N (xi , ac ), and let nci be the number of examples in this set. In Step 12 to Step 19, we calculate the score for each example and ask the oracle to label the example with the largest score. To be speciﬁc, for each class c, if we have not found any example from this class, we set the score of xi to be the maximum difference of nci and that of the neighboring points with similarity bigger than or equal to atc , where t is the iteration index. By querying the label of the example with the largest score, we are focusing on the regions where the underlying density changes the most. If this example is not from class c, we increase t by 1 and repeat; otherwise, we proceed to the next class. Notice that for a labeled example, any unlabeled example with global similarity bigger than or equal to the class speciﬁc similarity will not be selected in the future.

Justiﬁcation Global similarity matrix ∀c = 1, . . . , m, let Sc be the subset of S that consists of all the examples from class c, which are denoted Sc = {x1 (c), . . . , xKc (c)} ⊂ S. Let W c be the associated pair-wise similarity matrix, Kc × Kc , whose elements are deﬁned as in Equation 3.12. Notice that by setting σ to be the smallest distance to the K th nearest neighbor, we guarantee that for any example from a minority class, its pair-wise similarity with at least another example from the same minority class is reasonably Next, Kc large. c = c . Finally W deﬁne the diagonal matrix D c , Kc × Kc , where Dii k=1 ik deﬁne the normalized matrix W c = (Dc )−1/2 W c (Dc )−1/2 . Notice that W c is positive semi-deﬁnite, so its eigenvalues are non-negative, which are denoted λc1 ≥ λc2 ≥ . . . ≥ λcKc ≥ 0, and the corresponding eigenvectors are denoted uc1 , . . . , ucKc , s.t., uci = 1, i = 1, . . . , Kc . Furthermore, the largest eigenvalue λc1 is 1, with eigenvector uc1 ∝ (Dc )1/2 1Kc ×1 , where 1Kc ×1 is a vector of 1s. With respect to uc1 , we have the following lemma. Lemma 3. If σ changes with the number of examples n such that limn→∞ σ = 0 and limn→∞ n(σ)d = ∞, then as n goes to inﬁnity, ∀c = 1, . . . , m, uc1 · (uc1 )T converges in probability to Cc U c , where Cc is a constant, U c c = is a Kc × Kc matrix, its elements at the ith row and kth column Uik fc (xi (c)) × fc (xk (c)), and fc (xi (c)) is the probability density function of class c at xi (c).

57

3.3 Rare Category Detection for Graph Data

Algorithm 5 Graph-based Rare Category Detection (GRADE) Input: Unlabeled data set S, p1 , . . . , pm , α Output: The set I of selected examples and the set L of their labels 1: Let K = maxm c=2 n × pc . 2: For each example, calculate the distance between this example and its K th nearest neighbor. Set σ to be the minimum value of all such distances. 3: Construct the pair-wise similarity matrix W , n × n, where n is the number of examples, and ∀i, k = 1, . . . , n, Wik = exp(−

4: 5: 6: 7: 8: 9: 10: 11: 12: 13: 14: 15:

xi − xk 2 )I(i = k) 2σ 2

(3.12)

where I(·) is the indicator function. , i = Construct the diagonal matrix D, n × n, where Dii = nk=1 Wik 1, . . . , n. Calculate the normalized matrix W = D−1/2 W D −1/2 . Calculate the global similarity matrix A = (In×n −αW )−1 , where In×n is the n × n identity matrix. for c = 2 : m do Let Kc = npc . For each row of A, ﬁnd the (Kc )th largest element. Set ac to be the largest value of all such elements. ∀xi ∈ S, let N N (xi , ac ) = {x|x ∈ S, A(x, xi ) ≥ ac }, and nci = |N N (xi , ac )|, where A(x, xi ) is the corresponding element in A. end for for c = 2 : m do If class c has been discovered, continue. for t = 2 : n do For each xi that has been labeled yi , ∀xk ∈ S, if A(xi , xk ) ≥ ayi , max ac (nci − nck ). sk = −∞; for all the other examples, si = xk ∈N N (xi ,

t

)

Select and query the label of x = arg maxxi ∈S si . 17: If the label of x is equal to c, break; otherwise, mark the class that x belongs to as discovered. 18: end for 19: end for 16:

c Proof. As we have mentioned before, the ith element of uc1 : uc1 (i) ∝ Dii . On the other hand, as n goes to inﬁnity, Kc = npc goes to inﬁnity, and

58 c Dii √ (Kc −1)( 2πσ)d

3 Rare Category Detection

=

Kc

x (c)−x (c)2

k exp(− i 2σ 2 √ (Kc −1)( 2πσ)d

k=1,k=i

)

converges in probability to

E(δ(xi (c), x)) = δ(xi (c), x)fc (x)dx = fc (xi (c)), where δ(xi (c), x) is a delta function at x = xi (c). This is based on both the law of large numx (c)−x2

exp(− i 2 √ 2σ ( 2πσ)d

)

converges to δ(xi (c), x). Therefore, as n goes to inﬁnity, uc1 (i) is in proportion to fc (xi (c)), and uc1 · c (uc1 )T converges in probability to Cc U c , where Uik = fc (xi (c))fc (xk (c)).

bers and the fact that as σ goes to 0,

If we knew the class labels of all the examples in S, we could group the examples from the same class, and put the examples from the majority class at the end. To start with, suppose that if xi and xk are from different classes, Wik = 0. Then the normalized matrix W is block-diagonal, i.e., W = diag(W 2 , . . . , W m−1 , W 1 ). Therefore A = (In×n − αW )−1 is also block-diagonal, and it satisﬁes the following lemma. c 1 c c Lemma 4. If xi and xk both belong to class c, Aik = K l=1 1−αλcl ul (i)ul (k); otherwise, Aik = 0. 1 Proof. It is easy to see that the eigenvalues of A are 1−αλ c , c = 1, . . . , m, i i = 1, . . . , Kc , and the eigenvectors have the same value as uci if the corresponding example is from class c, and 0 otherwise. Reconstructing A based on eigen-decomposition, we get the above lemma.

If α is very close to 1, A can be approximated as follows. If xi and xk 1 both belong to class c, Aik ≈ 1−α uc1 (i)uc1 (k). According to Lemma 3, as n Cc fc (xi )fc (xk ). goes to inﬁnity, Aik converges in probability to 1−α Notice that uci = 1, c = 1, . . . , m, i = 1, . . . , Kc . In general, the absolute value of the elements of uc1 for the minority classes are much larger than that for the majority class since the majority class has far more examples than the minority classes. Therefore, if two examples are both from a minority class, their global similarity tends to be much larger than if they are both from the majority class. Compared with the pair-wise similarity matrix W , the global similarity matrix A is better suited for rare category detection. This is because if the minority class has a manifold structure, two examples on this manifold may be far away from each other in terms of Euclidean distance, so their pairwise similarity is very small; whereas their global similarity is large since it is roughly in proportion to the density of the minority class at both points. Fig.3.13a shows an example where the majority class (blue dots) has a uniform distribution, and the minority class (red balls) forms a 1-dimensional

3.3 Rare Category Detection for Graph Data

59

manifold. The black triangles at both ends of the manifold have a small pair-wise similarity. However, in terms of global similarity, they are quite similar, which matches our intuition. Furthermore, if we take the global similarity matrix as the pair-wise similarity matrix, and map all the points to the original feature space while preserving pair-wise similarity, the examples from the minority classes tend to form more compact clusters compared with the original feature representation (Fig.3.13b), whereas the probability density function of the majority class is still quite smooth, which makes the following querying process more effective. This is particularly beneﬁcial if the manifold structures of the minority classes are elongated, as shown in Fig.3.13a. Querying process Step 7 to Step 19 select examples to be labeled by the oracle. According to our discussion in the last subsubsection, if we reconstruct the features according to the global similarity matrix, the minority classes will form compact clusters and the probability density function of the majority class will be locally smooth. Generally speaking, the querying process selects examples from the regions where the local density changes the most, which have a high probability of coming from the minority classes. To be speciﬁc, as discussed before, if two examples are both from a minority class, their global similarity tends to be much larger than if they are both from the majority class. So the class speciﬁc similarity ac is likely to be determined by the examples from the minority classes. Furthermore, as n goes to inﬁnity, Aik is roughly in proportion to the density of class c at xi and xk if they both belong to minority class c. Therefore, if fc (xi ) is large, the global similarity between xi and the other examples from class c tends to be large, and nci is large accordingly. In other words, if we take the global similarity as the pair-wise similarity based on the new feature representation, nci is the number of neighbors within a ﬁxed distance. Therefore, nci is roughly in proportion to the local density at xi . For each class c, we calculate the score of each example xi , which is the maximum difference in the local density between xi and the other examples with global similarity bigger than or equal to atc . By querying the data point with the largest score, we focus on the regions where the local density changes the most, so we have a high probability of ﬁnding examples from the minority classes. Furthermore, by increasing the value of t, we gradually enlarge the size of the neighborhood. In this way, we are not only able to select points on the boundary of the minority classes, but also points in the interior, thus increasing our chance of ﬁnding the minority class examples. Finally, we make use of a simple feedback strategy: if an unlabeled example is quite similar to a labeled one, we preclude it from being selected

60

3 Rare Category Detection

in the future. In this way, we avoid wasting labeling effort on the minority classes that have been discovered already. Notice that the feedback strategy is orthogonal to the other components of our algorithm. Currently, we are exploring more effective feedback strategies. It should be mentioned that we do not make any assumption about the separability between the majority class and the minority classes, which is different from [FM06] and [PM04]. In fact, our algorithm works well when the support regions of the majority class and the minority classes overlap. The NNDB algorithm proposed in Section 3.1 can be seen as a special case of our algorithm for the binary case. If we use the pair-wise similarity matrix W instead of the global similarity matrix A, and the update of the neighborhood size is slightly modiﬁed in Step 15, our algorithm queries the same examples as NNDB. In the NNDB algorithm, it has been proven that under certain conditions, with a high probability, after a few iteration steps, NNDB queries at least one example whose probability of coming from the minority class is at least 13 . If the new feature representation based on the global similarity matrix satisﬁes these conditions, our algorithm shares the same theoretical properties as NNDB. In real applications, our algorithm is better than NNDB or MALICE (the counterpart of NNDB for multiple classes) since we make use of global similarity instead of pair-wise similarity, which makes the minority class examples more tightly clustered with the new feature representation. Application to Graph Data The proposed GRADE algorithm essentially makes use of the graph structure to help with rare category detection. Previously we have discussed its application on data with feature representations, which is based on the construction of the pair-wise similarity matrix using features. In this book, we mainly focus on this type of data. On the other hand, for graph data, it can be readily applied on the input graph. To be speciﬁc, given a graph G = (V, W ), where V = {v1 , . . . , vn } consists of all the vertices, W is the connectivity matrix, i.e., Wik is the edge weight if vi is connected with vk , and Wik =0 otherwise. W can be either binary or real-valued (non-negative). Notice that the elements of W denote the pair-wise similarity, which is similar to the pair-wise similarity matrix constructed in Step 3 of Alg. 5 for data with feature representation. To detect rare categories using Alg. 5, we input the graph G, p1 , . . . , pm and α. Then, we skip Step 1 to Step 3. All the other steps are the same as before. It is worth mentioning that in graph mining, researchers have developed

3.3 Rare Category Detection for Graph Data

61

algorithms for detecting dense subgraphs or communities [FLG00, KNRT03, GKT05]. If we want to use these approaches for rare category detection, it is labor-intensive to have the oracle label the whole subgraph. Conversely, the problem of picking representative vertices of the subgraphs for the oracle to label for detection purposes has not been addressed by existing work.

3.3.2

GRADE-LI Algorithm

In the GRADE algorithm, we need to input the proportions of all the classes. Notice that in practise, it is often difﬁcult to estimate the number of classes in the data set, not to mention the priors of different classes. However, it may be relatively easier to obtain an upper bound on the proportions of the minority classes of interest to us. In this subsection, we relax this requirement to produce the GRADE-LI algorithm, which only needs an upper bound p on the proportions of all the minority classes. Compared with GRADE, GRADE-LI is more suited for real applications. The GRADE-LI algorithm is summarized in Alg. 6. It works as follows. Step 1 to Step 3 construct the pair-wise similarity matrix. The only difference from the GRADE algorithm is that here we use the upper bound p to set the value of K. Step 4 to Step 6 calculate the global similarity matrix, which is the same as in the GRADE algorithm. Step 7 calculates the largest global similarity to the K th nearest neighbor and assigns it to a. Then in Step 8, for each example xi , we ﬁnd the number ni of its neighbors with global similarity bigger than or equal to a. The while loop in Step 9 is essentially the same as in the GRADE algorithm except that we are using a single similarity a instead of a set of class-speciﬁc similarities. If we only have one minority class, and p = p2 , GRADE and GRADE-LI produce the same result. If we have multiple classes, GRADE-LI requires less information than GRADE. The larger the number of minority classes, the greater the reduction in the amount of information needed as input. If the proportions of different minority classes do not vary a lot, a well represents the set of class speciﬁc similarities, and the performance of GRADE-LI is similar to GRADE. Furthermore, as we will show in the next subsection, GRADE-LI is quite robust to small perturbations in the upper bound p.

3.3.3

Experimental Results

In this subsection, we present the experimental results on both synthetic and real data sets to show the effectiveness of GRADE and GRADE-LI.

62

3 Rare Category Detection

2 1.5 1 0.5 0 −0.5 −1 −1.5 −2 −2

−1

0

1

2

(a) Original feature space: the two black triangles at both ends of the manifold have a small pair-wise similarity. 2 1.5 1 0.5 0 −0.5 −1 −1.5 −2 −2

−1

0

1

2

(b) Mapped feature space: the two black triangles are close to each other after we map the data to the feature space according to global similarity.

Figure 3.13: Synthetic data set: blue dots denote the majority class, and red balls denote the minority class.

63

3.3 Rare Category Detection for Graph Data

Algorithm 6 Graph-based Rare Category Detection with Less Information (GRADE-LI) Input: Unlabeled data set S, p, α Output: The set I of selected examples and the set L of their labels 1: Let K = n × p. 2: For each example, calculate the distance between this example and its K th nearest neighbor. Set σ to be the minimum value of all such distances. 3: Construct the pair-wise similarity matrix W , n × n, where n is the number of examples, and ∀i, k = 1, . . . , n, Wik = exp(−

4: 5: 6: 7: 8: 9: 10: 11:

xi − xk 2 )I(i = k) 2σ 2

Construct the diagonal matrix D, n × n, where Dii = nk=1 Wik ,i= 1, . . . , n. Calculate the normalized matrix W = D−1/2 W D −1/2 . Calculate the global similarity matrix A = (In×n − αW )−1 . For each row of A, ﬁnd the K th largest element. Set a to be the largest value of all such elements. ∀xi ∈ S, let N N (xi , a) = {x|x ∈ S, A(x, xi ) ≥ a}, and ni = |N N (xi , a)|. while not all the classes have been discovered do for t = 2 : n do For each xi that has been labeled yi , ∀xk ∈ S, if A(xi , xk ) ≥ a, max a (ni − nk ). sk = −∞; for all the other examples, si = xk ∈N N (xi , t )

Select and query the label of x = arg maxxi ∈S si . 13: Mark the class that x belongs to as discovered. 14: end for 15: end while 12:

Synthetic Data Sets Fig. 3.14 shows the result of applying GRADE on the synthetic data set in Fig. 3.13a. There are 1000 examples from the majority class, and only 20 examples from the minority class. Using random sampling, we need to label 51 examples to discover the minority class on average, whereas using the GRADE algorithm, we only need to label one example, denoted by the green ‘x’. Note that in this data set, we only have one minority class. If we run GRADE-LI with the prior of the minority class as input, we get exactly the

64

3 Rare Category Detection

same result as GRADE. 2 1.5 1 0.5 0 −0.5 −1 −1.5 −2 −2

−1

0

1

2

Figure 3.14: Synthetic data set: the example selected by GRADE is denoted by a green ‘x’. Fig. 3.15 shows another synthetic data set. It has one majority class (blue dots) and four minority classes (red balls), which form the four characters. The majority class has 1000 examples, whereas each minority class has 100 examples. In order to ﬁnd all the minority classes, using random sampling, we need to label 29 examples on average, whereas using the GRADE algorithm, we only need to label four examples, one from each minority class, which are denoted by green ‘x’s. Similar to GRADE, if we apply GRADE-LI on this data set with the exact upper bound, i.e., p = maxm c=2 pc , we need four label requests to ﬁnd all the minority classes. Based on the synthetic data set in Fig. 3.15, we gradually reduce the number of examples that form the character ‘I’, and compare the total number of label requests needed by GRADE and GRADE-LI in Fig. 3.16. As the number of examples from ‘I’ decreases, the proportions of different minority classes become more skewed, and the difference between GRADE and GRADE-LI becomes signiﬁcant. This matches our intuition since GRADELI only has access to an upper bound on the proportions of all the minority

65

3.3 Rare Category Detection for Graph Data

classes. Therefore, if the proportions of different minority classes are quite different, it may not be very good at discovering the smallest class. 8 6 4 2 0 −2 −4 −6 −8 −8

−6

−4

−2

0

2

4

6

8

Figure 3.15: Synthetic data set: the examples selected by GRADE are denoted by green ‘x’s. Real Data Sets In this subsubsection, we perform experiments on four real data sets, which are summarized in Table 3.7. Note that we have pre-processed the data so that each feature component has mean 0 and standard deviation 1. In the following experiments, we have compared GRADE and GRADE-LI with the following methods: MALICE, Interleave (the best method proposed in [PM04]) and random sampling (RS). Notice that the results for Interleave and RS are averaged over 100 runs. Fig. 3.17 to Fig. 3.20 show the comparison results on the 4 data sets. Note that for GRADE-LI, we use the exact upper bound as input. On the Ecoli data set, to discover all the classes, Interleave needs 41 label requests on average, MALICE needs 36 label requests, RS needs 43 label requests on average, GRADE needs 6 label requests, and GRADE-LI needs 32 label

66

3 Rare Category Detection

requests; on the Glass data set, to discover all the classes, Interleave needs 24 label requests on average, MALICE needs 18 label requests, RS needs 31 label requests on average, and both GRADE and GRADE-LI need 14 label requests; on the Abalone data set, to discover all the classes, Interleave needs 333 label requests on average, MALICE needs 179 label requests, RS needs 483 label requests on average, GRADE needs 149 label requests, and GRADE-LI needs 318 label requests; on the Shuttle data set, Interleave needs 140 label requests on average, MALICE needs 87 label requests, RS needs 512 label requests on average, GRADE needs 33 label requests, and GRADELI needs 36 label requests.

40

Number of Selected Examples

35 30 GRADE 25

GRADE−LI

20 15 10 5 0 10

20

30

40

50

60

70

80

90

100

Number of Examples in ’I’

Figure 3.16: Comparison between GRADE and GRADE-LI on the synthetic data set in Fig. 3.15. From these results, we have the following observations. First, on all the data sets, GRADE is much better than MALICE, which is the prior best method for rare category detection. Notice that both algorithms need the number of classes as well as the proportions of all the classes as input. Second, the performance of GRADE-LI is better than MALICE on all the data sets except the Abalone data set. The reason might be the following. In the Abalone data set, the proportion of the majority class is 16.50%, the propor-

3.3 Rare Category Detection for Graph Data

67

tion of the largest minority class is 15.18%, and the proportion of the smallest minority class is 0.34%. As we have shown with the synthetic data set in the last subsubsection, if the proportions of different minority classes do not vary a lot, which is the case for the other three data sets, the performance of GRADE-LI is similar to GRADE. On the other hand, if the proportions of different minority classes vary a lot, which is the case for the Abalone data set, the performance of GRADE-LI is worse than GRADE. It should be pointed out that compared with MALICE, GRADE-LI needs much less information: only an upper bound on the proportions of the minority classes is needed. The reduction in the prior knowledge of the data set is signiﬁcant especially when the number of classes in the data set is large, such as Abalone data set. The GRADE-LI algorithm needs an upper bound on the proportions of all the minority classes as input. Next we study the robustness of GRADE-LI with respect to this upper bound using the four real data sets. To this end, we add and subtract 15% from the exact upper bounds, and provide GRADE-LI with the perturbed upper bounds. In Fig. 3.21, we compare the following ﬁve methods in terms of the total number of label requests: MALICE, GRADE, GRADE-LI, GRADE-LI with p = 0.85 × maxm c=2 pc , and GRADE-LI with p = 1.15 × maxm p . c c=2 From Fig. 3.21, we can see that GRADE-LI is quite robust against small perturbations in the upper bounds. For example, with the Glass data set, to ﬁnd all the classes, using GRADE-LI, if p = maxm c=2 pc , we need 14 m label requests; if p = 0.85 × maxc=2 pc , we need 16 label requests; if p = 1.15 × maxm c=2 pc , we need 18 label requests.

3.3.4

Discussion

In this subsection, we discuss some implementation issues of GRADE and GRADE-LI. Pair-wise similarity Note that in Step 3 of Alg. 5 and Alg. 6, the Gaussian kernel is used to get the pair-wise similarity (Equation 3.12). In general, we can use the function ϕ(·, ·) to deﬁne W , which satisﬁesthe following conditions: ϕ(xi , xk ) = d k φ( xi −x φ(x)dx = 1, supx φ(x) < ∞, and σ ), ∀x ∈R , φ(x) ≥ 0, d limx→∞ φ(x) j=1 xj = 0, where xj is the j th feature component of x. In this case, it can be proven that Lemma 3 still holds. In practise, besides the above similarity functions, we can use the one that is best suited for our application, such as the cosine similarity for text data.

68

3 Rare Category Detection

Proportion of Classes Discovered

1 0.9 0.8 0.7 0.6 Interleave 0.5

MALICE RS

0.4

GRADE GRADE−LI

0.3 0.2 0.1 0

5

10

15

20

25

30

35

40

45

Number of Selected Examples

Figure 3.17: Comparison on Ecoli data set.

Proportion of Classes Discovered

1 0.9 0.8 0.7 0.6 Interleave 0.5

MALICE RS

0.4

GRADE GRADE−LI

0.3 0.2 0.1 0

5

10

15

20

25

30

Number of Selected Examples

Figure 3.18: Comparison on Glass data set.

35

69

3.3 Rare Category Detection for Graph Data

1

Proportion of Classes Discovered

0.9 0.8 0.7 Interleave

0.6

MALICE RS

0.5

GRADE 0.4

GRADE−LI

0.3 0.2 0.1 0 0

50

100

150

200

250

300

350

400

450

500

Number of Selected Examples

Figure 3.19: Comparison on Abalone data set.

Proportion of Classes Discovered

1 0.9 0.8 0.7 Interleave MALICE

0.6

RS 0.5

GRADE GRADE−LI

0.4 0.3 0.2 0.1 0

100

200

300

400

500

Number of Selected Examples

Figure 3.20: Comparison on Shuttle data set.

600

70

3 Rare Category Detection

Calculating the global similarity matrix The time complexity of both GRADE and GRADE-LI is dominated by the matrix inversion in calculating the global similarity matrix in Step 6 of Alg. 5 and Alg. 6, which is O(n3 ) with naive implementation, and O(n2.376 ) with Coppersmith and Winograd’s implementation [CW87]. To speed up this process, we can make use of one of the following two strategies. • The global similarity matrix A is deﬁned as (In×n − αW )−1 . Based

i i on Taylor’s expansion, we have A = In×n + ∞ i=1 α W . Therefore, we can use the following iterative process to get A, i.e., A0 = In×n , that A = A∞ , and we can Al = αW × Al−1 + In×n . It is easy to see i i approximate A with AL , i.e., A ≈ In×n + L i=1 α W . This strategy is best suited for the cases where W is a sparse matrix with z non-zero elements. In this case, with naive implementation, the complexity of calculating A is O(Lnz), which can be further improved with more sophisticated methods such as [YZ05].

• Another way to accelerate the calculation of the global similarity ma-

trix is by means of eigen-decomposition of W . Let W = ni=1 λi ui uTi , where λi is the ith largest eigenvalue, andui is the corresponding 1 eigenvector s.t., ui = 1. Therefore, A = ni=1 1−αλ ui uT . We can i i approximate W using eigenvalues, i.e., W ≈ ri=1 λi ui uTi . rthe r largest 1 In this way, A ≈ i=1 1−αλi ui uTi . Since we are only interested in the r largest eigenvalues of W , by making use of efﬁcient algorithms such as [LS96], the processing time will be greatly reduced.

Table 3.7: Properties of the data sets used. Data Set Ecoli [AN07] Glass [AN07] Abalone [AN07] Shuttle [AN07]

n 336 214 4177 4515

d 7 9 7 9

m 6 6 20 7

Largest Class 42.56% 35.51% 16.50% 75.53%

Smallest Class 2.68% 4.21% 0.34% 0.13%

Stopping criterion In GRADE-LI the querying process is stopped once we have at least one labeled example from each class. Note that the rest of GRADE-LI does not need the information of the number of classes to work. When this information is not available, we need to design a stopping criterion for the querying

71

3.3 Rare Category Detection for Graph Data

process. One possible criterion is to set a threshold on the number of consecutive uninformative label requests, which is the number of consecutive labeled examples whose classes have already been discovered, and stop the querying process once the threshold is reached. A very small threshold may result in the overlooking of certain minority classes; whereas a very large threshold may result in unaffordable labeling efforts. Therefore, we need to set a proper threshold according to speciﬁc applications. It should be pointed out that when the stopping criterion is not based on the number of classes in the data set, GRADE-LI may need to label more examples, but the number of label requests to discover all the classes is the same as reported in Subsection 3.3.3.

Number of Selected Examples

350 MALICE GRADE GRADE−LI −15% +15%

300

250

200

150

100

50

0

Ecoli

Glass

Abalone

Shuttle

Figure 3.21: Robustness study: -15% denotes the performance of GRADELI after we subtract 15% from the exact upper bounds; +15% denotes the performance GRADE-LI after we add 15% to the exact upper bounds. Picking the value of α As we have discussed in Subsection 3.3.1, if the pair-wise similarity between examples from different classes is 0, and the value of α is close to 1, the global similarity matrix A can be approximated using the eigenvectors that correspond to eigenvalue 1. In real applications, the pair-wise similarity between examples from different classes is rarely 0. In this case, if the

72

3 Rare Category Detection

value of α is close to 1, the resulting matrix A would be unreliable since the eigenvectors that correspond to eigenvalue 1 are not directly related to the probability density function of each class. Therefore, for the synthetic data sets, we set α = 0.8; whereas for the real data sets, we set α = 0.6.

3.4 Summary of Rare Category Detection In this chapter, we have discussed our work on rare category detection. Different from existing work, we target the challenging cases where the support regions of the majority and minority classes overlap with each other in the feature space. The overlapping phenomenon is observed in many real applications, such as ﬁnancial fraud detection and spam image detection. Yet it has not been well studied before. For data with feature representations, we propose the NNDB, ALICE, and MALICE algorithms, which need full prior information of the data set as input, including the number of classes and the proportions of different classes. We also propose the SEDER algorithm, which is prior-free. The basic idea of these algorithms is to select the examples from the regions where the density changes the most, since these examples have a high probability of coming from the minority classes. For graph data (relational data), we propose the GRADE algorithm, which needs full prior information of the data set as input, and the GRADE-LI algorithm, which only needs an upper bound on the proportions of the minority classes. The basic idea of these algorithms is to perform implicit feature transformation based on global similarity, which results in more compact clusters of the minority classes. Furthermore, we provide theoretical guarantees for these algorithms under the smoothness assumption of the majority classes and the compactness assumption of the minority classes. Based on the experimental results on both synthetic data sets and real data sets, we have the following conclusions. 1. Given full prior information about the data set, including the number of classes and the proportions of different classes, the proposed GRADE algorithm performs the best on our data sets, since it enjoys the beneﬁts of both the global similarity and the MALICE algorithm. 2. If we only know an upper bound on the proportions of the minority classes, but do not know the number of classes or the exact proportions of different classes, we can apply the GRADE-LI algorithm to detect the rare categories. 3. The GRADE-LI algorithm is robust to small perturbations in the upper

3.4 Summary of Rare Category Detection

73

bound. 4. The GRADE and GRADE-LI algorithms essentially make use of the graph structure to help with rare category detection. They can be applied to both data with feature representations and graph data. For data with feature representations, we ﬁrst need to construct a pair-wise similarity matrix based on the features. 5. In our experiments, the performance of GRADE is better than MALICE, and both of them are much better than existing methods, such as random sampling and Interleave (the best method proposed in [PM04]). However, the processing time of GRADE is more than MALICE since it needs to compute the global similarity for each pair of data points. 6. The MALICE algorithm is robust to small perturbations in the class priors. 7. If we do not have any prior information about the data set, we can apply the SEDER algorithm to detect the rare categories. In our experiments, if a data set is ‘moderately’ skewed, its performance is comparable to MALICE; if a data set is ‘highly’ skewed, its performance is not as good as MALICE. However, based on our experiments, the performance of SEDER is always better than random sampling.

3.4.1

Limitations

The limitations of the current work on rare category detection can be summarized as follows. 1. Violations of the assumptions. It is not clear how the performance of our algorithms will deteriorate if one of more of the assumptions is not satisﬁed. For example, the majority class may consist of several sub-classes of different sizes. One possible solution is to perform clustering on the original unlabeled data set so that each cluster only consists of one sub-class of the majority class and (probably) some minority class examples. Then we can use the techniques introduced in this chapter to discover the minority class examples from each cluster. 2. Computational issues. Some of the proposed algorithms do not scale well to very large data sets. For example, in MALICE, a major component is nearest neighbor search. This can be very time-consuming even with a well designed data structure such as kd-trees [Moo91]. Parallelizing these algorithms may be one way to address this problem. 3. Application to real world data sets. The current algorithms are evalu-

74

3 Rare Category Detection ated using benchmark data sets only. It is of great interest to test their performance on real world data sets, e.g., ﬁnancial fraud detection. Furthermore, we may be able to adapt the techniques introduced in this chapter to the particular problem at hand and incorporate domain expertise. 4. Other density estimators. In SEDER, specially designed exponential families are used to estimate the density as well as to calculate the scores for each data point. However, there is no systematic comparison between this density estimator and the large number of other methods on density estimation from both theoretical and practical point of view.

Chapter 4

Rare Category Characterization In Chapter 3, we have introduced various algorithms for rare category detection, which result in a set of labeled examples. Based on this labeled set, a natural follow-up step is rare category characterization, i.e., to characterize the minority classes in order to identify all the rare examples in the data set. For example, in Medicare fraud detection, once we have discovered a bogus claim for durable equipments (e.g., wheelchairs, breathing machines), we may want to ﬁnd the fraud patterns related to such equipments in order to prevent similar fraudulent claims in the future. To this end, in this chapter, we focus on rare category characterization, the second task in the supervised setting. Rare category characterization is to characterize the minority classes for the purpose of understanding and correctly classifying those classes. Here, our key observation is that the minority classes often exhibit compactness. That is, each minority class often forms a compact cluster. For example, fraudulent people often make multiple similar transactions to maximize their proﬁts [CPF06]. For rare diseases, the patients with the same type of rare disease often share similar genes or chromosomal abnormalities [EUR05]. In this chapter, we propose RACH by exploring such compactness for rare category characterization. The core of RACH is to represent the minority classes with a hyper-ball. We present the optimization framework as well as an effective algorithm to solve it. Furthermore, we show how RACH can be naturally kernelized. We also analyze the complexity of RACH. Finally, we justify the effectiveness of the proposed RACH by both theoretical analysis and empirical evaluations. J. He, Analysis of Rare Categories, Cognitive Technologies, DOI 10.1007/978-3-642-22813-1_4, © Springer-Verlag Berlin Heidelberg 2012

75

76

4 Rare Category Characterization The key content of this chapter can be summarized as follows. Problem Formulation. We formulate the problem of rare category characterization as an optimization problem, which takes into account both labeled and unlabeled examples, and imposes different constraints for different types of data; Algorithm Design. We design an effective algorithm to ﬁnd the solution of the optimization problem. It repeatedly converts the original problem into a convex optimization problem, and solves it in its dual form by a projected subgradient method, which is well justiﬁed theoretically.

The rest of this chapter is organized as follows. In Section 4.1, we propose the optimization framework to provide a compact representation for the minority class with justiﬁcation, followed by the conversion of this framework to the convex optimization problem as well as its dual form. Then we introduce the RACH algorithm to solve the dual problem with performance guarantees in Section 4.2, and the kernelized RACH algorithm in Section 4.3. Finally, following the experimental results presented in Section 4.4, we give a brief summary of rare category characterization in Section 4.5.

4.1 Optimization Framework In this section, we present our optimization framework, after we introduce additional notation, assumptions and the pre-processing step. Besides the notation introduced in Chapter 1, we also assume that there is only one majority class and one minority class in the data set, i.e., m = 2. (Multiple majority and minority classes can be converted into several binary problems.) Throughout this chapter, we will use calligraphic capital letters to denote sets. Let x1 , . . . , xn1 ∈ Rd denote the labeled examples from the majority class, which correspond to yi = 1, i = 1, . . . , n1 ; let xn1 +1 , . . . , xn1 +n2 ∈ Rd denote the labeled examples from the minority class, which correspond to yj = 2, j = n1 +1, . . . , n1 +n2 ; let xn1 +n2 +1 , . . . , xn ∈ Rd denote all the unlabeled examples. Here, n1 , n2 , and n denote the number of labeled examples from the majority class, the number of labeled examples from the minority class, and the total number of examples, both labeled and unlabeled. d is the dimensionality of the input space. Our goal is to identify with high precision and recall a list of unlabeled examples which are believed to come from the minority class. For rare category characterization, we make the following assumptions: the rare examples from the same minority class are very close to each other,

4.1 Optimization Framework

77

whereas the examples from the same majority class may be scattered in the feature space. This assumption is also used in [WXWC07][PM04], etc, when dealing with imbalanced data sets, either explicitly or implicitly. For example, in ﬁnancial fraud detection, fraud transactions of the same type are very similar to each other, whereas the normal transactions are more scattered in terms of their features compared with these fraud transactions. Furthermore, we also assume that the rare examples can be enclosed by a minimum-radius hyper-ball in the input space without including too many majority class examples. This seemingly rigorous assumption will become more ﬂexible when we use the high-dimensional feature space instead of the input space via the kernel trick in Section 4.3. With this assumption, we allow the support regions of the majority and minority classes to overlap with each other. In other words, our algorithm works in both the separable and non-separable cases, which makes it suitable for real applications.

4.1.1

Pre-processing: Filtering

In the unlabeled data, there might be some examples which are far away from the hyper-ball. These examples can be safely classiﬁed as majority class examples without affecting the performance of our classiﬁer. Therefore, we ﬁrst ﬁlter the unlabeled data to exclude such examples from the following optimization framework, and only focus on the examples that are close to the hyper-ball. The ﬁltering process is described in Alg. 7. It takes both the labeled and the unlabeled examples as input, and outputs a set of unlabeled examples which are close to the hyper-ball. Here, n − n1 − n2 is the number of unlabeled examples after the ﬁltering process. The algorithm works as follows. It ﬁrst estimates the center b of the hyper-ball using one-class SVM [SPST+ 01] or a single labeled example; then it estimates the proportion p of rare examples in the unlabeled data using the labeled data; ﬁnally, it calculates the distance threshold Dthre based on p, which is used to ﬁlter out the unlabeled examples far away from the hyper-ball. Notice that 3×Dthre is actually used to ﬁlter the unlabeled data. This is to ensure that we do not miss any rare example. We should point out that the ﬁltering process is orthogonal to the other parts of the proposed algorithm. In the remainder of this chapter, unlabeled data (unlabeled examples) refers to the examples output by the ﬁltering process.

4.1.2

Problem Formulations

Now, we are ready to give the problem formulations for rare category characterization. We ﬁrst give its original formulation and illustrate its intuitions.

78

4 Rare Category Characterization

Then, we present its convex approximation together with its dual form. Algorithm 7 Filtering Process for Rare Category Characterization Input: x1 , . . . , xn1 +n2 , xn1 +n2 +1 , . . . , xn Output: xn1 +n2 +1 , . . . , xn 1: if n2 > 1 then 2: Estimate the center b of the hyper-ball by one-class SVM [SPST+ 01], using all the labeled minority examples 3: else 4: Set the center b = xn1 +1 5: end if 6: for i = n1 + n2 + 1, . . . , n do 7: Calculate the distance di = xi − b 8: end for n2 9: Calculate p = n +n ; set Dthre as the (n − n1 − n2 ) × pth smallest value 1 2 among all di (i = n1 + n2 + 1, . . . , n ); 10: n = n1 + n2 11: for i = n1 + n2 + 1, . . . , n do 12: if di ≤ 3 · Dthre then 13: n = n + 1, xn = xi 14: end if 15: end for Original Formulation. To ﬁnd the center and radius of the minimumradius hyper-ball, we construct the following optimization framework, which is inspired by one-class SVM [SPST+ 01]. Problem 4.1 min

R2 ,b,α,β

2

R + C1

n1

α i + C2

n−n 1 −n2

i=1 2

βk

k=1

s.t., xi − b ≥ R2 − αi , i = 1, . . . , n1 αi ≥ 0, i = 1, . . . , n1 xj − b2 ≤ R2 , j = n1 + 1, . . . , n1 + n2 xk − b2 ≤ R2 + βk−n1 −n2 , k = n1 + n2 + 1, . . . , n βk−n1 −n2 ≥ 0, k = n1 + n2 + 1, . . . , n where R is the radius of the hyper-ball; b is the center of the hyper-ball; C1 and C2 are two positive constants that weight the three terms in the objective function; α and β correspond to the non-negative slack variables for the

79

4.1 Optimization Framework

labeled examples from the majority class and the unlabeled examples; αi and βk are the ith and k th component of α and β respectively. In Problem 4.1, we minimize the squared radius of the hyper-ball and a weighted combination of the slack variables. Furthermore, we have three types of constraints with respect to the training data. The ﬁrst type is for the labeled examples from the majority class, i.e., they should be outside the hyper-ball. Notice that these are not strict constraints, and the labeled examples from the majority class falling inside the hyper-ball correspond to positive slack variables αi . In this way, we allow the support regions of the majority and minority classes to overlap with each other. The second type is for the labeled examples from the minority class, i.e., they should be inside the hyper-ball. In contrast, these are strict constraints, and the hyper-ball should be large enough to enclose all the labeled rare examples. The last type is for the unlabeled examples, i.e., we want the hyper-ball to enclose as many unlabeled examples as possible. Different from the second type of constraints, these constraints are not strict, and the examples falling outside the hyper-ball correspond to positive slack variables βk . The intuition of this type of constraints is that after the ﬁltering process, the unlabeled examples are all in the neighborhood of the minority class. The support region of the minority class should have a higher density compared with the rest of the neighborhood. Therefore, we want the hyper-ball to enclose as many unlabeled examples as possible. Convex Approximation of Problem 4.1. Note that Problem 4.1 is difﬁcult to solve due to the ﬁrst type of constraints on the labeled examples from the majority class, which make this framework non-convex in the center b. To address this problem, we approximate these constraints based on the ﬁrstorder Taylor expansion around the current center ˜b, and have the following optimization problem, which is convex. Problem 4.2 (Convex Approximation of Problem 4.1) min

R2 ,b,α,β

2

R + C1

n1

α i + C2

n−n 1 −n2

i=1

βi

i=1 2

s.t., R2 − αi − xi 2 + ˜b + 2(xi − ˜b)T b ≤ 0, i = 1, . . . , n1 αi ≥ 0, i = 1, . . . , n1 xj − b2 ≤ R2 , j = n1 + 1, . . . , n1 + n2 xk − b2 ≤ R2 + βk−n1 −n2 , k = n1 + n2 + 1, . . . , n βk ≥ 0, k = 1, . . . , n − n1 − n2 Based on Problem 4.2, we ﬁnd the solution to Problem 4.1 in an iterative

80

4 Rare Category Characterization

way. To be speciﬁc, in each iteration step, we form Problem 4.2 based on the current estimate ˜b of the center, ﬁnd the optimal R2 , b, α and β, and then update Problem 4.2 based on the new center b. We stop the iteration when the solutions in two consecutive steps are very close to each other or when the maximum number of iteration steps is reached. Dual Problem for Problem 4.2. It is obvious that Problem 4.2 satisﬁes Slater’s condition [BV04]. Therefore, we solve this problem via the following dual problem. Problem 4.3 (Dual Problem for Problem 4.2) max λ

n

2

λj xj −

j=n1 +1

−

i=1

2

λi xi + n1

n1

j=n1 +1 λj xj − i=1 λi (xi n j=n1 +1 λj n1 n

s.t., 1 +

n

n1

λi =

i=1

λi ˜b2

i=1

− ˜b)2

λj

j=n1 +1

0 ≤ λi ≤ C1 , i = 1, . . . , n1 0 ≤ λj , j = n1 + 1, . . . , n1 + n2 0 ≤ λk ≤ C2 , k = n1 + n2 + 1, . . . , n where λ is the vector of Lagrange multipliers, λi , i = 1, . . . , n1 are associated with the constraints on the labeled examples from the majority class, λj , j = n1 + 1, . . . , n1 + n2 are associated with the constraints on the labeled examples from the minority class, and λk , k = n1 + n2 + 1, . . . , n are associated with the constraints on the unlabeled examples. Furthermore, based on the KKT conditions of Problem 4.2, the center b of the hyper-ball can be calculated as follows. n n1 ˜ j=n1 +1 λj xj − i=1 λi (xi − b) n b= (4.1) j=n1 +1 λj

4.2 Optimization Algorithm: RACH Here, we present the proposed optimization algorithm to solve Problem 4.1. The basic idea is as follow: after an initialization step; we will recursively formulate Problem 4.2 using the current estimate ˜b for the center of the hyper-ball; and then solve Problem 4.2 in its dual form (Problem 4.3) by a projected subgradient method.

4.2 Optimization Algorithm: RACH

4.2.1

81

Initialization Step

First, we need to initialize the center b of the hyper-ball and the Lagrange multipliers λ in Problem 4.3, which is summarized in Alg. 8. It takes as input both the labeled and the unlabeled examples (after the ﬁltering process), and outputs the initial estimates of the center b and the Lagrange multipliers λ. In Step 1, it initializes the center b and the radius R of the hyper-ball using one-class SVM [SPST+ 01] if we have more than one labeled example from the minority class; otherwise, it uses the only labeled rare example as the center b, and the smallest distance between this example and the nearest labeled example from the majority class as R. In Step 2, it initializes the Lagrange multipliers based on the KKT conditions of Problem 4.1. For a labeled example from the majority class, if its distance to the center b is bigger than R, λi = 0; if the distance is less than R, λi = C1 ; and if the distance is equal to R, we use C21 as the value for λi . For a labeled example from the minority class, if its distance to the center b is less than R, λj = 0; 2 otherwise, we use C1 +C as the value for λj . For an unlabeled example, if 2 its distance to the center b is less than R, λk = 0; if the distance is bigger than R, λk = C2 ; and if the distance is equal to R, we use C22 as the value for λk . Algorithm 8 Initialization for RACH Input: x1 , . . . , xn Output: initial estimates of b and λ 1: if n2 > 1 then 2: initialize the center b and the radius R of the hyper-ball using oneclass SVM [SPST+ 01] 3: else 4: set b = xn1 +1 , and set R as the smallest distance between xn1 +1 and the nearest labeled example from the majority class 5: end if 6: Initialize λ as follows. For 1 ≤ i ≤ n1 , if xi − b > R, λi = 0; if xi − b < R, λi = C1 ; if xi − b = R, λi = C21 For n1 + 1 ≤ j ≤ n1 + n2 , if xj − b < R, λj = 0; if 2 xj − b = R, λj = C1 +C 2 For n1 + n2 + 1 ≤ k ≤ n, if xk − b < R, λk = 0; if xk − b > R, λk = C2 ; if xk − b = R, λk = C22

82

4 Rare Category Characterization

4.2.2

Projected Subgradient Method for Problem 4.3

Projected subgradient methods minimize a convex function f (λ) subject to the constraint that λ ∈ X , where X is a convex set, by generating the sequence {λ(t) } via λ(t+1) = (λ(t) − τt ∇(t) ) X

is the (sub)gradient of f at λ(t) , τt is the step size, and X (x) = where arg miny {x−y : y ∈ X } is the Euclidean projection of x onto X . To solve 1 Next, we will foProblem 4.3, the gradient descent step is straight-forward. n 1 cus on the projection step, where X = {λ : 1+ i=1 λi = nj=n1 +1 λj , 0 ≤ λi ≤ C1 , i = 1, . . . , n1 ; 0 ≤ λj , j = n1 + 1, . . . , n1 + n2 ; 0 ≤ λk ≤ C2 , k = n1 + n2 + 1, . . . , n}. In the projection step, we consider the following optimization problem. Problem 4.4 (Projection Step of Problem 4.3) ∇(t)

1 min λ − v22 λ 2

s.t.,

n

ai λi = z, 0 ≤ λi ≤ εi

i=1

where ai (i = 1, . . . , n) denotes a set of constants which are either 1 or -1; z is a constant; v can be seen as the updated vector for λ based on gradient descent in each iteration step of the projected subgradient method, or λ(t) − τt ∇(t) ; and εi is the upper bound for λi . Without loss of generality, we assume that εi > 0, i = 1, . . . , n. For this optimization problem, deﬁne S+ = {i : 1 ≤ i ≤ n, ai = 1}, and S− = {i : 1 ≤ i ≤ n, ai = −1}. Before we give our optimization algorithm for Problem 4.4, we ﬁrst give the following lemma, which is the key to solving Problem 4.4.2 Lemma 5. Let λ be the optimal solution to Problem 4.4. Let s and t be two indices such that s, t ∈ S+ or s, t ∈ S− , and vs > vt . If λs = 0, then λt must be zero as well. On the other hand, let s and t be two indices such that s , t ∈ S+ or s , t ∈ S− , and vs − εs < vt − εt . If λs = εs , then λt must be εt as well. Proof. The Lagrange function of Problem 4.4: L(λ, θ, ζ, η) =

n n n 1 λ − v2 + θ( ai λi − z) − ζi λi − ηi (εi − λi ) 2 i=1

1

i=1

i=1

Note that in our case, we are maximizing a concave function in Problem 4.3, and gradient ascent is actually used in RACH. 2 Note that in [DSSSC08], the authors addressed a much simpler problem where ai = 1, and εi = ∞, i = 1, . . . , n.

4.2 Optimization Algorithm: RACH

83

where θ is a Lagrange multiplier associated with the equality constraint; ζ and η are two vectors of Lagrange multipliers associated with the inequality constraints whose elements are ζi and ηi respectively. Taking the partial derivative of L(λ, θ, ζ, η) with respect to λ and setting it to 0, we get λi = vi − ai θ + ζi − ηi

(4.2)

For the ﬁrst half of Lemma 1, suppose that s, t ∈ S+ . If λs = 0 and λt > 0, we have ζs ≥ 0, ηs = 0, ζt = 0 and ηt ≥ 0. Therefore, vs − θ + ζs = 0 and vt − θ − ηt > 0, which cannot be satisﬁed simultaneously since vs > vt . Therefore, if λs = 0, λt must be zero as well. Similar proof can be applied when s, t ∈ S− . For the second half of Lemma 1, suppose that s , t ∈ S+ . If λs = εs and λt < εt , we have ζs = 0, ηs ≥ 0, ζt ≥ 0 and ηt = 0. Therefore, λs −εs = vs −εs −θ−ηs = 0 and λt −εt = vt −εt −θ+ζt < 0, which cannot be satisﬁed simultaneously since vs −εs < vt −εt . Similar proof can be applied when s, t ∈ S− . Besides the vector v, deﬁne the vector v such that its ith element vi = vi − εi . Based on Lemma 5, for S+ (S− ), we can keep two lists: the ﬁrst list sorts the elements of v whose indices are in S+ (S− ) in ascending order, and only a top portion of the list corresponds to 0 in λ; the second list sorts the elements of v whose indices are in S+ (S− ) in descending order, and only a top portion of the list corresponds to the elements of λ that reach their upper bounds. For the remaining indices in S+ (S− ), their corresponding elements in λ are between 0 and the upper bound, and the Lagrange multipliers ζi = ηi = 0. Therefore, according to Equation 4.2, λi = vi − θ (λi = vi + θ). Finally, with respect to the value of θ, we have the following lemma. 1 2 3 , S+ and S+ Lemma 6. Let λ be the optimal solution to Problem 4.4. Let S+ denote subsets of S+ which correspond to the elements in λ that are equal to 0, equal 0 and the upper bound respectively. 2tothe3upper bound, and1 between 1 2 and S 3 denote subsets of S which S+ S+ = S+ . Let S− , S− S+ − − correspond to the elements in λ that are equal to 0, equal to the upper 2 3 bound, 1 and between 0 and the upper bound respectively. S− S− S− = S− . θ can be calculated as follows. 3 vk + 2 εj − 3 vk − 2 εj − z k∈S+ j∈S+ k∈S− j∈S− θ= (4.3) 3 3 |S+ | + |S− | 1 , S 2 , S 3 and S 1 , S 2 , S 3 , ∀i ∈ Proof. to the deﬁnition of S+ − − − According + 2 + 1 1 2 S+ S− , λi = 0, ζi ≥ 0, ηi = 0; ∀j ∈ S+ S− , λj = εj , ζj = 0, ηj ≥ 0;

84

4 Rare Category Characterization

3 3 3 S− , 0 < λk < εk , ζk = ηk = 0. Furthermore, ∀k ∈ S+ ∀k ∈ S+ , 3 λk = vk − θ; ∀k ∈ S− , λk = vk + θ. Therefore, z=

n

ai λi =

2 i∈S+

i=1

=

2 j∈S+

εj +

λi − 3 S+

(vk − θ) −

3 k∈S+

2 i∈S−

2 j∈S−

λi 3 S−

εj −

(vk + θ)

3 k∈S−

Solving this equation with respect to θ, we get Equation 4.3. Based on Lemma 5 and Lemma 6, to solve Problem 4.4, we gradually 1 2 1 2 increase the number of elements in S+ , S+ , S− and S− , calculate θ accordingly, and determine the value of λ which has the smallest value of 12 λ−v22 . Alg. 9 gives the details for solving Problem 4.3 in RACH. In Step 3 of Alg. 9, we calculate the gradient of the objective function in Problem 4.3 at the current value of λ and ˜b; then in Step 4, λ is updated via gradient ascent to obtain v. The remaining steps (Steps 5 - 16) are for the projection step (i.e., for solving Problem 4.4): in Step 5, we calculate the vector v using both v and the upper bounds; to calculate the projection of v 1 , S 2 , S 1 and S 2 , on X , in Step 7 to Step 15, we try different sizes for S+ + − − calculate θ and λ accordingly, and determine the projection of v based on the distance between v and w, where w is calculated based on the current sets 1 2 3 1 2 3 S+ , S+ , S+ , S− , S− and S− .

4.2.3

RACH for Problem 4.1

Algorithm. Now, we are ready to present the RACH algorithm (Alg. 10) to solve Problem 4.1. It is given the training data, the step size τ , C1 , C2 , and the number of iteration steps N1 , N2 . (Note that N2 is used in Alg. 9 in Step 3 of RACH.) The output is the unlabeled examples whose predicted class labels are class #2. RACH works as follows. First of all, it initializes the center b and the Lagrange multipliers λ using Alg. 8; then it repeatedly forms Problem 4.3 based on the current estimate of the center b, and applies Alg. 9 to solve it, which is based on the projected subgradient method; after solving Problem 4.3, the center b is updated using Equation 4.1; ﬁnally, we classify the unlabeled examples based on their corresponding Lagrange multipliers λk . The last step can be justiﬁed as follows. In Problem 4.1, for the unlabeled instances xk , k = n1 + n2 + 1, . . . , n, if xk − b < R, λk = 0, then yk = 2; if xk − b = R, 0 < λk < C2 , then yk = 2; if xk − b > R, λk = C2 , then yk = 1.

4.2 Optimization Algorithm: RACH

85

Algorithm 9 Projected Subgradient Method for Problem 4.3 Input: x1 , . . . , xn ; step size τ ; C1 , C2 ; N2 ; ˜b; initial estimate of λ Output: λ 1: Deﬁne S+ = {n1 + 1, . . . , n} and S− = {1, . . . , n1 } 2: for step = 1 to N2 do 3: Calculate ∇ as follows: if l = 1, . . . , n1 , ∇l = −xl 2 + ˜b2 + 2(

4: 5: 6: 7: 8:

9: 10:

11: 12:

λi (xi −˜b)T )(xl −˜b) ; if l = n1 + 1, . . . , n, ∇l = λ j=n1 +1 j n n T 1 λ (x −˜ n ( j=n +1 λj (xj )T − i=1 i i b) ) 1 · ( 2 j=n1 +1 λj (2xl − xj ) + ( n λ ) j=n +1 j

n j=n1 +1

λj (xj )T − n

n1 i=1

xl 2 − 1 n1 ˜ λ (x − b)) i=1 i i Update λ via gradient ascent to obtain: v = λ + τ ∇ Calculate v as follows: if i = 1, . . . , n1 , vi = vi − C1 ; if j = n1 +1, . . . , n1 +n2 , vj = −∞; if k = n1 +n2 +1, . . . , n, vk = vk −C2 Set D = ∞ for I1 = 1, I2 = 1, I3 = 1, I4 = 1 to I1 = |S+ | + 1, I2 = |S− | + 1, I3 = |S+ | + 1, I4 = |S− | + 1 do 1 ⊂ S denote the subset of indices in S whose correLet S+ + + sponding elements in v are no larger than the I1th largest element; 2 ⊂ S denote the subset of indices in S whose correspondlet S+ + + th ing elements in v are no smaller than the I smallest element 3 2 1 2 S+ = ∅ or S+ {n If S+ + 1, . . . , n + n }

= ∅, continue; 1 2 12 3 1 otherwise, S+ = S+ \(S+ S+ ) 1 ⊂ S denote the subset of indices in S whose correLet S− − − sponding elements in v are no larger than the I2th largest element; 2 ⊂ S denote the subset of indices in S whose correspondlet S− − − th ing elements in v are no smaller than the I smallest element 4 2 1 3 1 2) If S− S− = ∅, continue;otherwise, S− = S− \(S− S− Calculate θ =

3 k∈S+

vk −

2 2 3 vk +|S+ |C2 −|S− |C1 −1 k∈S− 3 |+|S 3 | |S+ −

1 1 2 S− , wi = 0; if i ∈ S+ Calculate w as follows: if i ∈ S+ , 2 3 3 , wi = C2 ; if i ∈ S− , wi = C1 ; if i ∈ S+ , wi = vi − θ; if i ∈ S− w i = vi + θ 14: If v − w < D, set λ = w and D = v − w. 15: end for 16: end for 13:

Concise Representation of the Minority Class. From Alg. 10, we can also compute the radius R of the hyber-ball, which is the maximum distance from the center b to xn1 +1 , . . . , xn whose Lagrange multipliers are less than

86

4 Rare Category Characterization

the corresponding upper bounds. The resulting hyber-ball (fully described by the center b and the radius R) provides a concise representation for the minority class. This representation can help us better understand the minority class. We can also use it to predict an unlabeled example as follows: if it is within the hyber-ball (i.e., its distance to the center b is less than R), we classify it as a rare example; otherwise, it belongs to the majority class. Furthermore, domain experts who have no knowledge in machine learning and data mining may be able to better understand the learning results based on this representation. Computational Complexity of RACH. It can be shown that the time complexity of RACH is O(N1 N2 (n − n1 )2 (n1 )2 ). In practice, we can reduce the running time in the following three ways. First, we ﬁnd that in our experiments RACH converges very quickly, often within a few tens of iterations. Second, in the applications that we are interested in, there are very few labeled examples from both the majority and the minority classes. A typical value for n1 is a few tens, and a typical value for n2 is less than 10. Finally, recall that in Section 4.1, we have applied Alg. 7 to ﬁlter out the unlabeled examples which are far away from the minority class. After this operation, only a small portion of the unlabeled data (typically less than 10%) is passed on to Alg. 10.

Algorithm 10 RACH: Rare Category Characterization Input: x1 , . . . , xn ; step size τ ; C1 , C2 ; N1 , N2 Output: unlabeled examples whose predicted class labels are 2 1: Initialize the hyper-ball center b and the Lagrange multipliers λ by Alg. 8 2: for step = 1 to N1 do 3: Update the Lagrange multipliers λ by Alg. 9 based on the current center b 4: Update the center b based on Equation 4.1 5: end for 6: for k = n1 + n2 + 1 to n do 7: if λk < C2 then 8: set yk = 2 9: else 10: set yk = 1 11: end if 12: end for

4.3 Kernelized RACH Algorithm

4.3

87

Kernelized RACH Algorithm

In this section, we brieﬂy introduce how to generalize RACH to the highdimensional feature space induced by kernels. The major beneﬁt of kernelizing the RACH algorithm is that, instead of enclosing the rare examples by a minimum-radius hyper-ball, we can now use more complex shapes, which make our algorithm more ﬂexible and may lead to more accurate classiﬁcation results. Compared with the original Alg. 10 which is designed for the input space, in the kernelized RACH algorithm, we only need to make the following changes. First, instead of directly calculating n the center b, we keep a set of coefﬁcients γi , i = 1, . . . , n such that b = i=1 γi xi . Therefore, Step 1 of Alg. 10 generates a set of initial coefﬁcients for b, and Step 4 updates the coefﬁcients of b based on Equation 4.1. In this way, b·x = ni=1 γi K(xi , x), and b2 = ni=1 nj=1 γi γj K(xi , xj ), where K(·, ·) is the kernel function. Next, notice that Alg. 10 and Alg. 9 are dependent on the examples only through the inner products or distances, which can be replaced by the kernel calculation.

4.4

Experimental Results

In this section, we present experimental results showing the effectiveness of RACH with the following synthetic data set as well as some real data sets. Fig. 4.1(a) shows a synthetic data set where the majority class has 3000 examples drawn from a Gaussian distribution, and the four minority classes correspond to four different shapes with 84, 150, 267, and 280 examples respectively. In this ﬁgure, the green circles represent labeled examples from the minority classes, and the red pluses represent labeled examples from the majority class. Here we construct four binary problems (the majority class vs. each minority class). Fig. 4.1(b) shows the classiﬁcation result where the green dots represent the rare examples, and the red dots represent the majority class examples. From this ﬁgure, we can see that almost all the rare examples are correctly identiﬁed except for a few points on the boundary. We also did experiments on seven real data sets, which are summarized in Table 4.1. For each data set, we constructed several binary problems consisting of one majority class and one minority class, and vary the percentage of labeled examples. For the sake of comparison, we also tested the following methods: (1) KNN (K-nearest neighbor); (2) Manifold-Ranking [ZWG+ 03]; (3) Under-Sampling; (4) TSVM [Joa99] with different costs for the examples from different classes; (5) SVM-Perf [Joa05]. We used the RBF kernel in RACH. All the parameters are tuned by cross validation.

88

4 Rare Category Characterization

(a) Input data set: the green circles represent labeled examples from the minority classes, and the red pluses represent labeled examples from the majority class.

(b) Classiﬁcation results: the examples in green are classiﬁed as rare examples, and the examples in red are classiﬁed as majority class examples.

Figure 4.1: Rare category characterization on the synthetic data set.

89

4.4 Experimental Results 1 0.9 0.8

F−score

0.7 0.6 0.5 0.4 KNN Manifold−Ranking Under−Sampling TSVM RACH SVM−Perf

0.3 0.2 0.1 0 0.05

0.1

0.15

0.2

0.25

Label Percentage (a) Class 1 (689 examples) vs. Class 14 (67 examples) 1 0.98 0.96

F−score

0.94 0.92 0.9 0.88 0.86

KNN Manifold−Ranking Under−Sampling TSVM RACH SVM−Perf

0.84 0.82 0.8 0.05

0.1

0.15

0.2

Label Percentage (b) Class 2 (634 examples) vs. Class 4 (57 examples)

Figure 4.2: Results on Abalone data set

0.25

90

4 Rare Category Characterization 1

0.9

F−score

0.8 KNN Manifold−Ranking Under−Sampling TSVM RACH SVM−Perf

0.7

0.6

0.5

0.4

0.05

0.1

0.15

0.2

0.25

Label Percentage (a) Class 1 (143 examples) vs. Class 2 (77 examples) 1 0.9 0.8

F−score

0.7 0.6 0.5 0.4 KNN Manifold−Ranking Under−Sampling TSVM RACH SVM−Perf

0.3 0.2 0.1 0 0.05

0.1

0.15

0.2

Label Percentage (b) Class 2 (77 examples) vs. Class 3 (52 examples)

Figure 4.3: Results on Ecoli data set

0.25

91

4.4 Experimental Results 0.9 0.8 0.7

F−score

0.6 0.5 0.4 0.3 KNN Manifold−Ranking Under−Sampling TSVM RACH SVM−Perf

0.2 0.1 0 0.05

0.1

0.15

0.2

0.25

Label Percentage (a) Class 1 (70 examples) vs. Class 3 (17 examples)

1 0.9 0.8

F−score

0.7 KNN Manifold−Ranking Under−Sampling TSVM RACH SVM−Perf

0.6 0.5 0.4 0.3 0.2 0.1 0 0.05

0.1

0.15

0.2

Label Percentage (b) Class 1 (70 examples) vs. Class 5 (9 examples)

Figure 4.4: Results on Glass data set

0.25

92

4 Rare Category Characterization 1 0.95 0.9

F−score

0.85 0.8 0.75 0.7 0.65

KNN Manifold−Ranking Under−Sampling TSVM RACH SVM−Perf

0.6 0.55 0.5 0.05

0.1

0.15

0.2

0.25

Label Percentage (a) Class 1 (463 examples) vs. Class 6 (44 examples) 0.9 0.8 0.7

F−score

0.6 KNN Manifold−Ranking Under−Sampling TSVM RACH SVM−Perf

0.5 0.4 0.3 0.2 0.1 0 0.05

0.1

0.15

0.2

Label Percentage (b) Class 3 (244 examples) vs. Class 8 (30 examples)

Figure 4.5: Results on Yeast data set

0.25

93

4.4 Experimental Results 1 0.9 0.8

F−score

0.7 0.6 0.5 0.4 KNN Manifold−Ranking Under−Sampling TSVM RACH SVM−Perf

0.3 0.2 0.1 0 0.05

0.1

0.15

0.2

0.25

Label Percentage (a) Class 2 (329 examples) vs. Class 4 (88 examples) 1 0.9 0.8

F−score

0.7 KNN Manifold−Ranking Under−Sampling TSVM RACH SVM−Perf

0.6 0.5 0.4 0.3 0.2 0.1 0 0.05

0.1

0.15

0.2

Label Percentage (b) Class 2 (329 examples) vs. Class 5 (115 examples)

Figure 4.6: Results on Page Blocks data set

0.25

94

4 Rare Category Characterization 1 0.9 0.8

F−score

0.7 KNN Manifold−Ranking Under−Sampling TSVM RACH SVM−Perf

0.6 0.5 0.4 0.3 0.2 0.1 0 0.05

0.1

0.15

0.2

0.25

Label Percentage (a) Class 3 (132 examples) vs. Class 2 (37 examples) 1 0.95 0.9 0.85

KNN Manifold−Ranking Under−Sampling TSVM RACH SVM−Perf

F−score

0.8 0.75 0.7 0.65 0.6 0.55 0.5 0.05

0.1

0.15

0.2

Label Percentage

(b) Class 2 (37 examples) vs. Class 7 (11 examples)

Figure 4.7: Results on Shuttle data set

0.25

95

4.4 Experimental Results 0.45 0.4 0.35

F−score

0.3 0.25 0.2 0.15 KNN Manifold−Ranking Under−Sampling TSVM RACH SVM−Perf

0.1 0.05 0 0.05

0.1

0.15

0.2

0.25

Label Percentage (a) comp (4852 examples) vs. misc.forsale (964 examples) 0.45 0.4 0.35

F−score

0.3 0.25 0.2 0.15 KNN Manifold−Ranking Under−Sampling TSVM RACH SVM−Perf

0.1 0.05 0 0.05

0.1

0.15

0.2

0.25

Label Percentage

(b) rec (3968 examples) vs. comp.os.ms-windows.misc (963 examples)

Figure 4.8: Results on 20 Newsgroups data set

96

4 Rare Category Characterization

The comparison results in terms of the F-score (harmonic mean of precision and recall) of the minority class are shown in Fig. 4.2 to Fig. 4.8 where the x-axis is the percentage of labeled examples. Under each ﬁgure, the numbers outside the brackets are the class indices included in the binary problem, and the numbers inside the brackets are the number of examples from each class. From these ﬁgures, we can see that the performance of RACH is consistently better than the other methods across all the data sets, especially when the percentage of labeled examples is small, which is the case we are most interested in. In particular, the performance of RACH is better than SVMPerf in most cases, although the latter directly optimizes the F-score. This might be due to the fact that the objective function of SVM-Perf is only an upper bound of the training loss regularized by the L2 norm of the weight vector. Also, SVM-Perf is designed for general classiﬁcation problems; and it might ignore the skewness and the compactness properties of the minority class. On the other hand, the performance of the other methods varies a lot across the different data sets. For example, in Fig. 4.2(a), the performance of KNN is only worse than RACH; whereas in Fig. 4.3(b), the performance of KNN is worse than TSVM, and as the percentage of labeled examples increases, KNN performs less well than Under-Sampling. Table 4.1: Summary of the data sets Data Set No. of Examples No. of Features Data set No. of Examples No. of Features

Abalone 4177 7 Page Blocks 5473 10

Ecoli 336 7 Shuttle 4515 9

Glass Yeast 214 1484 9 8 20 Newsgroups 18774 61188

4.5 Summary of Rare Category Characterization In this chapter, we have discussed our work on rare category characterization. It follows the task of rare category detection, and makes use of labeled data from all the classes to ﬁnd a compact representation for the minority classes. The goal is to identify all (or nearly all) the rare examples in the data set with high precision and recall. Different from the large amount of work on imbalanced classiﬁcation, we only focus on the rare examples, and we are able to generate a compact representation for the minority classes for the sake of interpretability.

4.5 Summary of Rare Category Characterization

97

In our algorithm, the basic idea is to enclose the rare examples with a minimum-radius hyper-ball based on the clustering property of the minority classes. We formulate this idea as an optimization problem and present the effective optimization algorithm RACH to ﬁnd its solution. In RACH, we repeatedly (1) convert the original problem into a convex optimization problem, and (2) solve it in its dual form by a projected subgradient method. Furthermore, we generalize RACH to the high-dimensional feature space induced by kernels. Experimental results demonstrate the effectiveness of the proposed RACH.

4.5.1

Limitations

The limitations of the current work on rare category characterization can be summarized as follows. 1. Interactions with domain experts. In RACH, the compact representations for the minority classes are used to help domain experts better understand the learning results. However, in our current work, we did not perform experiments to validate the effectiveness of these representations. 2. Application to data sets with multiple minority classes. The current version of RACH can only be applied to the binary setting, i.e., there is one majority class and one minority class in the data set. When multiple minority classes are present in the data set, we may ﬁrst perform clustering on the labeled rare examples, and then apply RACH for each cluster, which corresponds to a single minority class. This generalization may greatly improve its applicability to real world problems. 3. Computational issues. The proposed RACH algorithm does not scale well to very large data sets (see the analysis on its computational complexity in Subsection 4.2.3). This may be addressed by parallelism and simple yet effective heuristics, such as localized analysis.

Chapter 5

Unsupervised Rare Category Analysis In this chapter, we focus on unsupervised rare category analysis, i.e., no label information is available in the learning process, and address the following two problems: (1) rare category selection, i.e., selecting a set of examples which are likely to come from the minority class; (2) feature selection, i.e., selecting the features that are relevant to identify the minority class. The key observation is that the above two tasks are correlated with each other. On one hand, the analysis of the minority class examples helps us identify the relevant features; on the other hand, the identiﬁcation of the relevant features is crucial to the selection of the minority class examples. Therefore, we propose to jointly deal with the two tasks so that they can beneﬁt from each other. To this end, we formulate the problem as a well-justiﬁed optimization framework, which co-selects the relevant features and examples from the minority class. Furthermore, we design an effective search procedure based on the augmented Lagrangian method. The basic idea is to alternately ﬁnd the relevant features and the minority class examples. Finally, we demonstrate the performance of the proposed method by experimental results on some benchmark data sets. The key content of this chapter can be summarized as follows. Problem Deﬁnition. To the best of our knowledge, we are the ﬁrst to address the two important tasks in unsupervised rare category analysis, and we propose to jointly deal with them; Problem Formulation. We design an optimization framework for the co-selection of features and examples, which is well justiﬁed theoretically; J. He, Analysis of Rare Categories, Cognitive Technologies, DOI 10.1007/978-3-642-22813-1_5, © Springer-Verlag Berlin Heidelberg 2012

99

100

5 Unsupervised Rare Category Analysis Search Procedure. We develop an effective algorithm to solve the optimization problem which is based on the augmented Lagrangian method.

The rest of this chapter is organized as follows. In Section 5.1, we present the optimization framework with theoretical justiﬁcation. Section 5.2 introduces the algorithm for solving the optimization problem. Experimental results are given in Section 5.3. Finally, we conclude this chapter with a brief summary in Section 5.4.

5.1 Optimization Framework In this chapter, we focus on the binary case, i.e., one majority class and one minority class, and our goal is to (1) select a set of examples which are likely to come from the minority class, and (2) identify the features relevant to this minority class. In this section, we formulate this problem as an optimization framework, and provide some theoretical justiﬁcation.

5.1.1

Additional Notation

In this chapter, we are dealing with the binary case, i.e., m = 2. Therefore, p1 = 1 − p2 . Furthermore, of the d features, only dr features are relevant to the minority class. In other words, the examples from the minority class have very similar values on those features, and their values on the other features may be quite diverse. For the sake of simplicity, assume that the dr features are independent of each other. Therefore, the examples from the minority class are tightly clustered in the dr -dimensional subspace spanned by the relevant features, which we call the relevant subspace. Let Sdr denote the set of all dr -dimensional subspaces of Rd , and let Smin denote the relevant subspace, Smin ∈ Sdr . Let f (x) denote the probability density function (pdf) of the data in Rd , i.e., f (x) = (1 − p2 )f1 (x) + p2 f2 (x), where f1 (x) and f2 (x) are the pdf of the majority and minority classes in Rd respectively. Given feature subspace S ∈ Sdr and x ∈ Rd , (S) let x(S) denote the projection of x on S, and f (S) (x(S) ), f1 (x(S) ) and (S) f2 (x(S) ) denote the projection of f (x), f1 (x) and f2 (x) on S respectively. To co-select the minority class examples and the relevant features, we deﬁne two vectors: a ∈ Rn and b ∈ Rd . Let ai and bj denote the ith and j th elements of a and b respectively. ai = 1 if xi is from the minority class, and 0 otherwise; bj = 1 if the j th feature is relevant to the minority class, and 0 otherwise.

101

5.1 Optimization Framework

5.1.2

Objective Function

Given the prior p2 of the minority class and the number of relevant features dr , we hope to ﬁnd np2 data points whose corresponding ai = 1, and dr features whose corresponding bj = 1. Intuitively, the np2 points should form a compact cluster in the relevant subspace, and due to the characteristic of the minority class, this cluster should be more compact than any other np2 data points in any dr -dimensional subspace. To be more strict, we have the following optimization problem. Problem 5.1

min f (a, b) =

n n d 1 ai ak ( bj (xji − xjk )2 ) np2 i=1 k=1

s.t.,

n

j=1

ai = np2 , ai = 0, 1

i=1 d

bj = dr , bj = 0, 1

j=1

In the objective function f (a, b), dj=1 bj (xji − xjk )2 is the squared distance between xi and xk in the subspace Sb spanned by the features with non-zero bj . This squared distance contributes to f (a, b) if and only if both ai and ak are equal to 1. Given a set of np2 points, deﬁne the set distance of every data point to be the sum of the squared distances between this point and all the points within this set. Therefore, by solving this optimization problem, we aim to ﬁnd a set of np2 points and dr features such that the average set distance of these points to this set in the corresponding subspace Sb is the minimum. Problem 5.1 can be easily applied to the case where either a or b is known, and we want to solve for the other vector. To be speciﬁc, if a is known, i.e., we know the examples that belong to the minority class, and we want to ﬁnd the dr -dimensional subspace where the minority class can be best characterized, we can use the same objective function f (a, b), and solve for b using the minority class examples. Similarly, if b is known, i.e., we know which features are relevant to the minority class, and we want to ﬁnd the examples from the minority class, we can also use f (a, b), and solve for a in the subspace Sb spanned by the relevant features.

102

5.1.3

5 Unsupervised Rare Category Analysis

Justiﬁcation

The optimization problem we introduced in the last subsection is intuitively reasonable. Next, we look at it from a theoretical perspective. ∀S ∈ Sdr , deﬁne function ψ S as follows. ∀S ∈ Sdr , x ∈ Rd , let 1 S (S) ψ (x ) = minDnp2 ⊂D,|Dnp2 |=np2 np2 y∈Dnp x(S) − y (S) 2 = np1 2 2 np2 (S) (i) 2 (i) th nearest neighbor of x(S) x − z , where z denotes the i i=1 x(S) x(S) (S)

(S)

within x1 , . . . , xn , i.e., ψ S (x(S) ) is the average squared distance between x(S) and its np2 nearest neighbors. Furthermore, deﬁne function φS as follows. φS (x(S) ) = E(ψ S (x(S) )). Here, the expectation is with respect to (i) zx(S) , i = 1, . . . , np2 . Based on the above deﬁnitions, we have the following theorem. Theorem 7. If 1. In Smin , the support region of the minority class is within hyper-ball B of radius r; 2. The support region of f in any dr -dimensional subspace is bounded, i.e., maxS∈Sdr maxx,y∈Rd ,f (S) (x(S) )>0,f (S) (y(S) )>0 x(S) − y (S) = α < +∞; 3. The density of the majority class in hyper-ball B is non-zero, i.e., (S ) miny∈Rd ,y (Smin ) ∈B (1 − p2 )f1 min (y (Smin ) ) = f0 > 0; 4. The function value of φS is big enough if the projection of the data point in the dr -dimensional subspace S is not within B, i.e., S (S) ) − 4r2 > β > 0; minS∈Sdr ,x∈Rd ,x(S) ∈B / φ (x 5. The number of examples is sufﬁciently large, i.e., n ≥ max{ 2(VB1f0 )2 8

· log 2δ , 4pα2 β 4 log Cddr

2

2Cddr δ

}, where VB is the volume of hyper-ball B, and

is the number of d choose dr ; then with probability at least 1 − δ, in the solution to Problem 5.1, the subspace Sb spanned by the features with bj = 1 is the relevant subspace Smin , and the data points with ai = 1 are within B. Proof. The basic idea of the proof is to show that if the selected feature subspace Sb is NOT Smin , or at least one point in the set of np2 points is outside B in Smin , we can always use Smin to ﬁnd another set of np2 points such that all the points are within B, and its objective function is smaller than the original set. To do this, ﬁrst, notice that according to condition (3), the expected proportion of data points falling inside B, E( nnB ) ≥ p2 + VB f0 , where nB denotes the number of points within B. Second, according to

103

5.1 Optimization Framework (i)

condition (2), ∀x ∈ D, Pr[0 ≤ x(S) − zx(s) 2 ≤ α2 ] = 1, i = 1, . . . , np2 . Therefore, nB < p2 or ∃x ∈ D, ∃S ∈ Sdr , s.t., ψ S (x(S) ) < φS (x(S) ) − β] n nB < p2 ] + Pr[∃x ∈ D, ∃S ∈ Sdr , s.t., ψ S (x(S) ) < φS (x(S) ) − β] ≤ Pr[ n nB nB − E( ) < −VB f0 ] + nCddr Pr[ψ S (x(S) ) < φS (x(S) ) − β] ≤ Pr[ n n nB nB − E( ) < −VB f0 ] + nCddr (np +1) Pr[ψ S (x(S) ) ≤ Pr[ n n z 2

Pr[

x(S)

(S)

< φ (x S

)−

(np +1) (np +1) β|zx(S)2 ]dPr[zx(S)2 ]

≤ exp(−2n(VB f0 )2 ) + nCddr exp(−

2np2 β 2 ) α4

where Cddr is an upper bound on the number of subspaces in Sdr , and the last inequality is based on Hoeffding’s inequality1 . 2 Let exp(−2n(VB f0 )2 ) ≤ 2δ , and nCddr exp(− 2npα24β ) ≤ 2δ , we get n ≥ 8

1 2(VB f0 )2

2Cddr δ .

log 2δ , and n ≥ 4pα2 β 4 log 2 examples n is sufﬁciently large, i.e., n ≥ max{

In other words, if the number of

2Cddr 2 α8 1 log log , } 2(VB f0 )2 δ 4p22 β 4 δ

then with probability at least 1 − δ, there are at least np2 points within hyperball B, and ∀x ∈ D, ∀S ∈ Sdr , ψ S (x(S) ) ≥ φS (x(S) ) − β. Furthermore, according to condition (4), ∀x ∈ D, ∀S ∈ Sdr , if x(S) ∈ / B, ψ S (x(S) ) > 4r 2 . (S ) Notice that ∀a, ∀b, f (a, b) ≥ i:ai =1 ψ Sb (xi b ). On the other hand, if Sb = Smin , and the points with ai = 1 are within B in Smin , then f (a, b) < 4np2 r2 . This is because the squared distance between any two points within B in Smin is no bigger than 4r2 . Given a and b, if Sb is not Smin , we can design a and b in such a way that Sb is Smin , and the points that correspond to ai = 1 are within B in Smin . We can always ﬁnd such a vector a since we have shown that there (S ) are at least np2 points within B. Therefore, f (a, b) ≥ i:ai =1 ψ Sb (xi b ) > 4np2 r2 > f (a , b ). On the other hand, if Sb is Smin , but at least one point with ai = 1 is outside B, we can design a and b in such a way that b = b, and a replaces the points with ai = 1 that is outside B with 1

(np +1)

Note that given zx(S)2

, ψ S (x(S) ) can be seen as the average of np2 independent items.

104

5 Unsupervised Rare Category Analysis

some points within B that are different from existing points in a. For the sake of simplicity, assume that only xt is outside B. Therefore, f (a, b) = (Smin ) (S ) (S ) (S ) 1 −xk min 2 + np2 2 ni=1 ai xi min −xt min 2 i=t k=t ai ak xi np2 (S ) (S ) (S ) ≥ np1 2 i=t k=t ai ak xi min −xk min 2 +2ψ Smin (xt min ) > np1 2 i=t (Smin ) (S ) − xk min 2 + 8r2 ≥ f (a , b ). The above derivation can k=t ai ak xi be easily generalized to the case where more than one point with ai = 1 are outside B. Therefore, in the solution to Problem 5.1, Sb is the relevant subspace Smin , and the data points with ai = 1 are within B. The conditions of Theorem 7 are straightforward except condition (4). According to this condition, ∀S ∈ Sdr , if x(S) ∈ / B and y (Smin ) ∈ B, S (S) Smin (Smin ) (y ) by at least β when there are at least φ (x ) is bigger than φ np2 points within B in Smin . Therefore, this condition can be roughly interpreted as follows. The density around x(S) is smaller than the density around y (Smin ) such that the expected average squared distance between x(S) and its np2 nearest neighbors is much larger than that between y (Smin ) and its np2 nearest neighbors. In this way, assuming the other conditions in Theorem 7 are also satisﬁed, with high probability, we can identify the relevant subspace and pick the examples within B according to a. It should be pointed out that if we want to select np2 points from the minority class, picking them from hyper-ball B is the best we can hope for. In this way, each selected example has a certain probability of coming from the minority class. On the other hand, if some selected points are outside B, their probability of coming from the minority class is 0.

5.2 Partial Augmented Lagrangian Method In this section, we introduce the Partial Augmented Lagrangian Method (PALM) to effectively solve Problem 4.1. In our method, we alternate the optimization of a and b, i.e., given the current estimate of a, we solve for b that leads to the minimum value of f (a, b); given the current estimate of b, we solve for a that decreases the value of f (a, b) as much as possible. To be speciﬁc, f (a, b) can be rewritten as f (a, b) = dj=1 bj ni=1 nk=1 j j 2 1 np2 ai ak (xi − xk ) . Therefore, given a, we can solve for b as follows. For each feature j, calculate its score saj = np1 2 ni=1 nk=1 ai ak (xji − xjk )2 . Then ﬁnd the dr features with the smallest scores, and set their corresponding bj = 1 and all others to 0. It is easy to show that this vector b minimizes f (a, b) given a. On the other hand, given b, solving for a is not straight-

105

5.2 Partial Augmented Lagrangian Method

forward, since f (a, b) is not a convex function of a. In this chapter, we ﬁrst relax the constraints on a: instead of requiring that ai be binary, we require that ai ∈ [0, 1], i.e., we solve the following optimization problem on a: Problem 5.2 n n d 1 ai ak ( bj (xji − xjk )2 ) np2

min gb (a) =

i=1 k=1

s.t.,

n

j=1

ai = np2 , ai ∈ [0, 1]

i=1

Next we use the augmented Lagrangian method [NW99] to solve Problem 5.2 in an iterative way. The reason for using the augmented Lagrangian method is the following: it is a combination of Lagrangian and quadratic penalty methods; the addition of the penalty terms to the Lagrangian function does not alter the stationary point of the Lagrangian function, and can help damp oscillations and improve convergence. Here, we deﬁne the following augmented Lagrangian function n n d 2n+1 1 j j 2 ai ak ( bj (xi − xk ) ) − λi di (a) LA (a, λ, σ) = np2 i=1 k=1

+

σ 2

2n+1

j=1

i=1

d2i (a)

(5.1)

i=1

where λi , i = 1, . . . , 2n + 1 are the Lagrange multipliers, σ is a positive parameter, and di (a), i = 1, . . . , 2n + 1 are a set of functions deﬁned as follows. c (a) if i ≤ 1 or c (a) ≤ i i λi otherwise σ n ai − np2 = 0 c1 (a) =

di (a) =

λi σ

i=1

cj+1 (a) = aj ≥ 0, 1 ≤ j ≤ n ck+n+1 (a) = 1 − ak ≥ 0, 1 ≤ k ≤ n Here ci (a), i = 1, . . . , 2n + 1, denote the original constraints on a, both equality and inequality, and di (a) are truncated versions of ci (a), i.e., di (a) is equal to ci (a) if and only if the corresponding constraint is active or nearactive; it is ﬁxed at λσi otherwise.

106

5 Unsupervised Rare Category Analysis

We minimize LA (a, λ, σ) based on Algorithm 4.20 in [MNT04]. Putting together the optimization of a and b, we have the Partial Augmented Lagrangian Method, which is presented in Alg. 11. Algorithm 11 Partial Augmented Lagrangian Method (PALM) Input: Initial values of λ and σ: λ0 and σ0 , stepmax Output: a and b 1: Initialize a and b 2: λ = λ0 , σ = σ0 , Kprev = d(a)∞ 3: for step = 1 to stepmax do 4: a := arg mina LA (a, λ, σ), K := d(a)∞ K 5: if K ≤ prev then 2 6: for i = 1 to 2n + 1 do 7: λi := λi − σdi (a) 8: end for 9: Kprev := min(K, Kprev ) 10: else 11: σ := 2 × σ 12: end if 13: for j = 1 to d do 14: Calculate the score for the j th feature saj = j j 2 1 n n i=1 k=1 ai ak (xi − xk ) np2 15: end for 16: Pick dr features with the smallest scores, and set their corresponding bj to 1 and all others to 0 17: end for The algorithm works as follows. Given the initial values λ0 and σ0 of λ and σ, and the maximum number of iteration steps stepmax , it will output vectors a and b that correspond to a local minimum of f (a, b). In Step 1, we initialize a and b. Next, in Step 2, we assign λ and σ to their initial values, and calculate Kprev , which is the maximum absolute value of all the di (a) functions, i = 1, . . . , 2n + 1. Then Step 4 to Step 16 are repeated stepmax times. In Step 4, we minimize the augmented Lagrangian function with respect to a, given the current estimates of λ, σ, and b. To be speciﬁc, we use gradient descent to update a, and gradually decrease the step size until convergence. Once we have obtained an updated estimate of a, we calculate K, which is the maximum absolute value of the current di (a) functions. If the value of K is less than half of Kprev , then we update the Lagrange

5.3 Experimental Results

107

multipliers using the formula in Step 7, which is called the steepest ascent formula in [MNT04]. Furthermore, we update Kprev using the smaller value of K and Kprev . Otherwise, if the value K is bigger than half of Kprev , we double the value of σ. Next, we update the value of b based on the current estimate of a. To be speciﬁc, for each feature, we calculate its score based on the formula in Step 14. Then in Step 16, we pick the dr features with the smallest scores, and set the corresponding bj to 1, which minimizes f (a, b) given a. In our experiments, the algorithm always converges around 20 iteration steps, so we set stepmax = 30. Notice that the vectors a and b generated by PALM correspond to a local minimum of f (a, b). To improve its performance, we can run PALM with different initializations of a and b in Step 1 of Alg. 11, and pick the best values of a and b that correspond to the smallest f (a, b). The vectors a and b can be interpreted as follows. For b, its dr non-zero elements correspond to the relevant features. For a, ideally the minority class examples should correspond to ai = 1. However, this may not be the case in practice. Therefore, we rank the elements of a from large to small, and hope to ﬁnd all the minority class examples in the ﬁrst dr positions of the ranked list.

5.3

Experimental Results

In this section, we demonstrate the performance of PALM from the following perspectives: (1) the quality of rare category selection; (2) the quality of feature selection; (3) the beneﬁt of co-selecting features and instances. In our experiments, we retrieve the minority class examples from the ranked list generated by different methods, and use the following performance measures: (1) the precision-scope curve, i.e., the percentage of the minority class examples when a certain number of examples are retrieved, such as 10% × np2 , . . . , 100% × np2 ; (2) the recall-scope curve, i.e., the percentage of the minority class examples when a certain number of MINORITY class examples are retrieved, such as 10% × np2 , . . . , 100% × np2 .

5.3.1

Synthetic Data Sets

To demonstrate the performance of PALM, we ﬁrst use a simple synthetic data set shown in Fig. 5.1. In this ﬁgure, there are 1000 examples from the majority class, denoted by blue dots, which are uniformly distributed in the feature space, and only 10 examples from the minority class, denoted by red balls, whose features on Z are uniformly distributed. Of the three features,

108

5 Unsupervised Rare Category Analysis

only two features (X and Y ) are relevant to the minority class, i.e., the minority class examples have very similar values on these features; and one feature (Z) is irrelevant to the minority class, i.e., the minority class examples spread out on this feature. Using PALM, given the number of minority class examples and the number of relevant features, we are able to identify the relevant features, with the corresponding bj = 1. Of the 10 examples with the largest ai values, nine examples are from the minority class, and the remaining minority class example has the 11th largest ai value.

Z

0.5

0

−0.5 0.5

0 0.5

Y

−0.5

0 −0.5

X

Figure 5.1: Synthetic data set: there are 1000 examples from the majority class, denoted by blue dots, and only 10 examples from the minority class, denoted by red balls. Next we test the precision of the selected features of PALM using synthetic data sets with different prior p2 . Fig. 5.2, Fig. 5.3, and Fig. 5.4 show the comparison results of PALM with the Laplacian score method [HCN05], feature variance method (selecting the features with the largest variance), CRO [KC07], and random sampling. The x-axis is the proportion of irrelevant features, and the y-axis is the precision of the selected features. From these results, we can see that PALM is much better than the other four methods especially when the prior p2 is small. As for the Laplacian score method, although it is comparable with PALM for large p2 , its performance quickly deteriorates as p2 decreases (e.g., Fig. 5.2a and b), which is the case we are interested in for rare category analysis.

109

5.3 Experimental Results

PALM Laplacian Variance CRO Random

1

Precision

0.8

0.6

0.4

0.2

0 0

0.2

0.4

0.6

0.8

1

Proportion of Irrelevant Features (a) p2 = 0.01

1 PALM Laplacian Variance CRO Random

Precision

0.8

0.6

0.4

0.2

0 0

0.2

0.4

0.6

0.8

1

Proportion of Irrelevant Features (b) p2 = 0.015

Figure 5.2: Precision of selected features on synthetic data sets (part 1).

110

5 Unsupervised Rare Category Analysis

1

Precision

0.8

0.6

0.4

0.2

0 0

PALM Laplacian Variance CRO Random 0.2

0.4

0.6

0.8

1

Proportion of Irrelevant Features (a) p2 = 0.02

1

Precision

0.8

0.6

0.4

0.2

0 0

PALM Laplacian Variance CRO Random 0.2

0.4

0.6

0.8

1

Proportion of Irrelevant Features (b) p2 = 0 05 .

Figure 5.3: Precision of selected features on synthetic data sets (part 2).

111

5.3 Experimental Results

1

Precision

0.8

0.6

0.4

0.2

0 0

PALM Laplacian Variance CRO Random 0.2

0.4

0.6

0.8

1

Proportion of Irrelevant Features (a) p2 = 0.1

1

Precision

0.8

0.6

0.4

0.2

0 0

PALM Laplacian Variance CRO Random 0.2

0.4

0.6

0.8

1

Proportion of Irrelevant Features (b) p2 = 0.2

Figure 5.4: Precision of selected features on synthetic data sets (part 3).

112

5.3.2

5 Unsupervised Rare Category Analysis

Real Data Sets

In this subsection, we test the performance of PALM on rare category selection. To the best of our knowledge, there are no existing methods for this task. Therefore, we have designed the following methods for the sake of comparison. 1. Random: randomly rank all the examples, and select the ﬁrst np2 points from the ranked list as the minority class examples. 2. NNDB-based: calculate the score of each example using NNDB [HC07]. Note that no feedback from the labeling oracle is available, so the scores are not updated. 3. Interleave-based: calculate the score of each example using the Interleave principle [PM04]. Similar to the NNDB-based method, the scores of the examples are not updated in this method. 4. PALM-full: assume that all the features are relevant to the minority class, i.e., bj = 1, j = 1, . . . , d, and run PALM with dr = d. Note that the NNDB-based method and the Interleave-based method are both derived from rare category detection methods. Here we use four real data sets, which are summarized in Table 5.1. In this chapter, we focus on binary problems, i.e., there is only one majority class and one minority class in the data set. Therefore, for each data set, we construct several subproblems as follows. We combine the examples from two different classes into a smaller binary data set, using the larger class as the majority class, the smaller class as the minority class, and test the different methods on these binary data sets. For PALM, we tune the number of relevant features dr without any label information. For each data set, we present the results on two binary subproblems, which are shown in Fig. 5.5 to Fig. 5.12. On the other binary subproblems, similar results are observed and therefore omitted for space. In these ﬁgures, the left ﬁgure shows precision vs. scope, and the right ﬁgure shows recall vs. scope. On all the data sets, PALM performs the best: the precision and recall sometimes reach 100%, such as Fig. 5.10 and Fig. 5.11. As for the other methods (Interleave-based, NNDB-based, and PALM-full), their performance depends on different data sets, and none of them is consistently better than Random. Comparing with Random, Interleave-based, and NNDBbased, we can see that PALM does a better job at selecting the minority class examples; comparing with PALM-full, we can see that the features selected by PALM indeed help improve the performance of rare category selection.

113

5.3 Experimental Results

1 0.9

Precision

0.8

PALM PALM−full Random NNDB−based Interleave−based

0.7 0.6 0.5 0.4 0.3 0.2

20

40

60

80

100

80

100

Scope (a)

1 0.9 0.8

PALM PALM−full Random NNDB−based Interleave−based

Recall

0.7 0.6 0.5 0.4 0.3 0.2

20

40

60

Scope (b)

Figure 5.5: Abalone data set: class 1 vs. class 7, p2 = 0.362, 4 features selected by PALM.

114

5 Unsupervised Rare Category Analysis

1.1 1 0.9

Precision

0.8

PALM PALM−full Random NNDB−based Interleave−based

0.7 0.6 0.5 0.4 0.3 0.2

20

40

60

80

100

80

100

Scope (a)

1.1 1 0.9

Recall

0.8

PALM PALM−full Random NNDB−based Interleave−based

0.7 0.6 0.5 0.4 0.3 0.2

20

40

60

Scope (b)

Figure 5.6: Abalone data set: class 2 vs. class 7, p2 = 0.381, 4 features selected by PALM.

115

5.3 Experimental Results

0.9 0.8 0.7

PALM PALM−full Random NNDB−based Interleave−based

Precision

0.6 0.5 0.4 0.3 0.2 0.1 0 −0.1

20

40

60

80

100

80

100

Scope (a)

1 0.9 0.8

Recall

0.7

PALM PALM−full Random NNDB−based Interleave−based

0.6 0.5 0.4 0.3 0.2 0.1 0

20

40

60

Scope (b)

Figure 5.7: Ecoli data set: class 1 vs. class 4, p2 = 0.197, 3 features selected by PALM.

116

5 Unsupervised Rare Category Analysis

PALM PALM−full Random NNDB−based Interleave−based

1

Precision

0.8

0.6

0.4

0.2

0

20

40

60

80

100

Scope (a)

1

PALM PALM−full Random NNDB−based Interleave−based

0.9 0.8

Recall

0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

20

40

60

80

100

Scope (b)

Figure 5.8: Ecoli data set: class 2 vs. class 4, p2 = 0.313, 4 features selected by PALM.

117

5.3 Experimental Results

PALM PALM−full Random NNDB−based Interleave−based

1

Precision

0.8 0.6 0.4 0.2 0

20

40

60

80

100

Scope (a)

PALM PALM−full Random NNDB−based Interleave−based

1

Recall

0.8 0.6 0.4 0.2 0 20

40

60

80

100

Scope (b)

Figure 5.9: Glass data set: class 1 vs. class 3, p2 = 0.195, 2 features selected by PALM.

118

5 Unsupervised Rare Category Analysis

PALM PALM−full Random NNDB−based Interleave−based

1

Precision

0.8

0.6

0.4

0.2

0 20

40

60

80

100

Scope (a)

1 PALM PALM−full Random NNDB−based Interleave−based

0.9 0.8

Recall

0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

20

40

60

80

100

Scope (b)

Figure 5.10: Glass data set: class 2 vs. class 3, p2 = 0.183, 3 features selected by PALM.

119

5.3 Experimental Results 1 0.9 0.8

Precision

0.7

PALM PALM−full Random NNDB−based Interleave−based

0.6 0.5 0.4 0.3 0.2 0.1 0

20

40

60

80

100

Scope (a)

1 0.9 0.8

Precision

0.7

PALM PALM−full Random NNDB−based Interleave−based

0.6 0.5 0.4 0.3 0.2 0.1 0

20

40

60

80

100

Scope (b)

Figure 5.11: Yeast data set: class 2 vs. class 6, p2 = 0.093, 2 features selected by PALM.

120

5 Unsupervised Rare Category Analysis 1 PALM PALM−full Random NNDB−based Interleave−based

0.9 0.8

Precision

0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

20

40

60

80

100

Scope (a)

1 PALM PALM−full Random NNDB−based Interleave−based

0.9 0.8

Precision

0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

20

40

60

80

100

Scope (b)

Figure 5.12: Yeast data set: class 2 vs. class 9, p2 = 0.055, 3 features selected by PALM.

5.4 Summary of Unsupervised Rare Category Analysis

121

Notice that in some ﬁgures (Fig. 5.5b, Fig. 5.6b, Fig. 5.7b, Fig. 5.9b, and Fig. 5.10b), at the end of the recall curves, the different methods seem to overlap with each other. This is because with no supervision, it is sometimes difﬁcult to retrieve all the examples from the minority class, and the last example from the minority class tends to appear towards the end of the ranked list. Therefore, the recall value at 100%np2 is often close to the prior of the minority class in the data set. Finally, we test the performance of PALM when there are small perturbations in the number of relevant features. To this end, we run PALM with dr increased by 1 (PALM+1) and decreased by 1 (PALM-1) respectively, and compare their performance with PALM and PALM-full in Fig. 5.13. From this ﬁgure, we can see that PALM is quite robust against small perturbations in dr in most cases (Abalone, Ecoli, and Glass), and in all the cases, the performance of PALM+1 and PALM-1 is better than PALM-full (i.e., PALM without feature selection). Table 5.1: Properties of the data sets [AN07] used. Data Set Ecoli Glass Abalone Yeast

5.4

n 336 214 4177 1484

d 7 9 7 8

Largest Class 42.56% 35.51% 16.50% 31.20%

Smallest Class 2.68% 4.21% 0.34% 1.68%

Summary of Unsupervised Rare Category Analysis

In this chapter, we have discussed our work on unsupervised rare category analysis. Different from the supervised settings, in the unsupervised settings, we do not have access to the labeling oracle. Under certain assumptions, we are able to address the following two problems: (1) rare category selection, and (2) feature selection. Here, our key observation is that jointly dealing with the two tasks beneﬁts both of them. To this end, we propose an optimization framework, which is well justiﬁed theoretically. To solve this optimization problem, we design the Partial Augmented Lagrangian Method (PALM), which alternately ﬁnds the relevant features and the minority class examples. The effectiveness of PALM is demonstrated by experimental results.

122

5 Unsupervised Rare Category Analysis

1

PALM−1 PALM PALM+1 PALM−full

0.9 0.8

Precision

0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

Abalone(2 vs.7) Ecoli(2 vs.4) Glass(1 vs.3) Yeast(2 vs.6)

(a) 1

PALM−1 PALM PALM+1 PALM−full

0.9 0.8

Recall

0.7 0.6 0.5 0.4 0.3 0.2 0.1

0 Abalone(2 vs. 7) Ecoli(2 vs.4) Glass(1 vs.3) Yeast(2 vs.6)

(b)

Figure 5.13: Perturbation on the number of relevant features.

5.4 Summary of Unsupervised Rare Category Analysis

5.4.1

123

Limitations

The limitations of the current work on unsupervised rare category analysis can be summarized as follows. 1. Violations of the assumptions. It is not clearly how the performance of PALM will deteriorate if one of more of the assumptions is not satisﬁed. For example, the support region of the minority class is elongated. 2. Application to data sets with multiple minority classes. The current version of PALM can only be applied to the binary setting, i.e., there is one majority class and one minority class in the data set. PALM can be generalized to multiple majority and minority classes in the data set by targeting the minority classes one at a time. However, no experiments have been performed in this respect.

Chapter 6

Conclusion and Future Directions Rare categories are of key importance in many real applications: although the occurrence of such examples is rare, their impact is signiﬁcant. Applications of rare category analysis include: ﬁnancial fraud detection, Medicare fraud detection, network intrusion detection, astronomy, spam image detection and health care. In this book, for the sake of analyzing rare categories, we make the following two assumptions: (1) smoothness assumption for the majority classes, and (2) compactness assumption for the minority classes. Notice that we do not assume the majority and minority classes are separable / near-separable from each other in the feature space, which is assumed by most existing work on rare category analysis. In other words, we target the more challenging case where the support regions of the majority and minority classes overlap with each other in the feature space (although some proposed algorithms work well in both the separable and non-separable cases), since many real applications exhibit the overlapping phenomenon. In this book, we have focused on both supervised and unsupervised rare category analysis. In particular, the following three tasks are addressed. 1. Rare category detection: discovering at least one example from each minority class with the fewest label requests to the labeling oracle. For data with feature representations, we propose the NNDB, ALICE, and MALICE algorithms that need full prior information of the data set as input, including the number of classes and the proportions of different classes; we also propose the SEDER algorithm, which is prior-free. For graph data (relational data), we propose the GRADE algorithm that needs full prior information as input; we also generalize this alJ. He, Analysis of Rare Categories, Cognitive Technologies, DOI 10.1007/978-3-642-22813-1_6, © Springer-Verlag Berlin Heidelberg 2012

125

126

6 Conclusion and Future Directions gorithm to produce the GRADE-LI algorithm that only needs an upper bound on the proportions of all the minority classes. For each of these algorithms, we provide theoretical justiﬁcations as well as empirical evaluations, showing their effectiveness on our data sets.

2. Rare category characterization: given labeled examples from all the classes, ﬁnding a compact representation for the minority classes in order to identify all the rare examples with high precision and recall. To this end, we propose to enclose the rare examples with a minimum-radius hyper-ball based on the clustering property of the minority classes. This idea is further formulated as an optimization problem, and we design the RACH algorithm to ﬁnd its solution. There are two key components in RACH: (1) converting the original problem into a convex optimization problem, and (2) solving it in its dual form by a projected subgradient method. RACH can be easily kernelized. Experimental results demonstrate the effectiveness of RACH. Furthermore, with the compact representation, we are able to let domain experts (who have no knowledge in machine learning and data mining) better understand the learning results. 3. Unsupervised rare category analysis: selecting the examples that are likely to come from the minority classes and selecting the features relevant to the minority classes in an unsupervised fashion. We propose to co-select the rare examples and the relevant features, which beneﬁts both tasks. To this end, we design an optimization framework, which is well justiﬁed theoretically. To solve this optimization problem, we propose the Partial Augmented Lagrangian Method (PALM), which alternately ﬁnds the relevant features and the minority class examples. The effectiveness of PALM is demonstrated by experimental results. Rare category analysis can be extended in multiple dimensions. For example, 1. Understanding the dynamics of rare categories. To be speciﬁc, how does a rare category emerge and evolve over time? Take emerging research topics as an example. A new research topic may start from a single paper exploring a new research direction, which can be followed by more researchers working in the same area, forming a minority class. Finally, it may become a major research topic. Understanding the dynamics of rare categories can help us gain deeper insights into these categories and provide tools for predicting their occurrence. To address this problem, as a ﬁrst step, we may want to monitor the number of papers on a new research topic over time to see if there is any

6 Conclusion and Future Directions

127

change in the distribution once the research topic has become a minority class. Then we may use the above observation to predict if a new research topic will eventually become a major research topic or gradually disappear. These techniques can also be used in disease evolution to monitor possible variations of a certain disease over time. 2. Understanding complex fraud. In many real world problems, the fraud patterns may be more complex than a single fraudulent transaction or a bogus claim. For example, in eBay, the fraudsters and accomplices form a bipartite core. The fraudulent identities are eventually used to carry out the actual non-delivery fraud, while the accomplices exist only to help the fraudsters by boosting their feedback rating [CPF06]. Such bipartite cores are a tell-tale sign of a popular fraud scheme, which has resulted in a total loss in the order of millions. To discover these fraud patterns, we need to focus on the subgraph level. As a ﬁrst step, we may apply the GRADE or GRADE-LI algorithms proposed in Section 3.3 on the whole transaction graph to discover the complex fraud. Then we can incorporate the bipartite nature of the subgraphs into the algorithms. For example, we may try to modify the global similarity so that vertices in the same bipartite graph tend to have larger global similarity. 3. Transfer learning for rare category analysis. In this aspect, we may focus on both inter-domain and intra-domain transfer learning. In the ﬁrst case, the goal is to leverage the information of rare categories in a source domain to help us understand the rare categories in the target domain. Here, the major challenge is the lack of label information related to the rare categories in the target domain. For example, compared with ﬁnancial fraud, Medicare fraud is less studied, but the different areas may share common fraud patterns. To address this problem, ﬁrst of all, in order to leverage the label information from the source domain, we need to build some connection between the rare categories in the source domain and the target domain. Do they have similar distributions in the feature space? Are certain parameters shared by these rare categories across different domains? Do they evolve over time in a similar way? Then, based on the shared properties of these rare categories, we can effectively transfer the label information from the source domain to the target domain. In the second case, we work in a single domain, and the goal is to make use of known rare categories to help us detect new rare categories. Here, the major challenge is the lack of label information related to

128

6 Conclusion and Future Directions the new rare categories. For example, in disease diagnosis, studying patients with known ﬂu types may help us detect new ﬂu variants. As before, to address this problem, we ﬁrst need to ﬁnd the shared properties of the rare categories, both known and unknown. Then we may have a better idea of the support regions for the new rare categories. By sampling in these regions, we may be able to reduce the number of label requests from the labeling oracle.

Bibliography [AN07]

A. Asuncion and D.J. Newman. UCI machine learning repository, 2007.

[Ang87]

Dana Angluin. Queries and concept learning. Machine Learning, 2(4):319–342, 1987.

[Ang01]

Dana Angluin. Queries revisited. In ALT, pages 12–31, 2001.

[AY01]

Charu C. Aggarwal and Philip S. Yu. Outlier detection for high dimensional data. In SIGMOD Conference, pages 37–46, 2001.

[BBL06]

Maria-Florina Balcan, Alina Beygelzimer, and John Langford. Agnostic active learning. In ICML, pages 65–72, 2006.

[BEYM05] R. Bekkerman, R. El-Yaniv, and A. McCallum. Multi-way distributional clustering via pairwise interactions. In ICML, pages 41–48, New York, NY, USA, 2005. ACM. [BKA+ 06]

Stephen Bay, Krishna Kumaraswamy, Markus G. Anderle, Rohit Kumar, and David M. Steier. Large scale detection of irregularities in accounting data. In ICDM, pages 75–86, 2006.

[BV04]

Stephen Boyd and Lieven Vandenberghe. Convex optimization. Cambridge University Press, 2004.

[BWJ01]

Daniel Barbar´a, Ningning Wu, and Sushil Jajodia. Detecting novel network intrusions using Bayes estimators. In Proceedings of the First SIAM Conference on Data Mining, April 2001.

[CAL92]

David Cohn, Les Atlas, and Richard Ladner. Improving generalization with active learning, 1992.

[CBHK02] Nitesh V. Chawla, Kevin W. Bowyer, Lawrence O. Hall, and W. Philip Kegelmeyer. SMOTE: Synthetic minority oversampling technique. J. Artif. Intell. Res. (JAIR), 16:321–357, 2002. J. He, Analysis of Rare Categories, Cognitive Technologies, DOI 10.1007/978-3-642-22813-1, © Springer-Verlag Berlin Heidelberg 2012

129

130

Bibliography

[CBK09]

Varun Chandola, Arindam Banerjee, and Vipin Kumar. Anomaly detection: A survey. ACM Comput. Surv., 41(3), 2009.

[CGJ96]

David A. Cohn, Zoubin Ghahramani, and Michael I. Jordan. Active learning with statistical models. Journal of Artiﬁcial Intelligence Research, 4:129–145, 1996.

[Cha09]

N.V. Chawla. Mining when classes are imbalanced, rare events matter more, and errors have costs attached. In SDM, 2009.

[CLHB03]

Nitesh V. Chawla, Aleksandar Lazarevic, Lawrence O. Hall, and Kevin W. Bowyer. SMOTEBoost: Improving prediction of the minority class in boosting. In PKDD, pages 107–119, 2003.

[CPF06]

Duen Horng Chau, Shashank Pandit, and Christos Faloutsos. Detecting fraudulent personalities in networks of online auctioneers. In PKDD, pages 103–114, 2006.

[CW87]

D. Coppersmith and S. Winograd. Matrix multiplication via arithmetic progressions. In STOC, pages 1–6, New York, NY, USA, 1987. ACM.

[Das05]

Sanjoy Dasgupta. Coarse sample complexity bounds for active learning. In NIPS, 2005.

[DB00]

Jennifer G. Dy and Carla E. Brodley. Feature subset selection and order identiﬁcation for unsupervised learning. In ICML, pages 247–254, 2000.

[DCSL02]

Manoranjan Dash, Kiseok Choi, Peter Scheuermann, and Huan Liu. Feature selection for clustering - a ﬁlter solution. In ICDM, pages 115–122, 2002.

[DE95]

Ido Dagan and Sean P. Engelson. Committee-based sampling for training probabilistic classiﬁers. In In Proceedings of the Twelfth International Conference on Machine Learning, pages 150–157. Morgan Kaufmann, 1995.

[DGBK07] Haimonti Dutta, Chris Giannella, Kirk D. Borne, and Hillol Kargupta. Distributed top-k outlier detection from astronomy catalogs using the DEMAC system. In SDM, 2007. [DH08]

Sanjoy Dasgupta and Daniel Hsu. Hierarchical sampling for active learning. In ICML, pages 208–215, 2008.

[DHS00]

R. O. Duda, P. E. Hart, and D. G. Stork. Pattern classiﬁcation.

131

Bibliography Wiley-Interscience Publication, 2000. [DMM03]

I.S. Dhillon, S. Mallela, and D.S. Modha. Informationtheoretic co-clustering. In KDD, pages 89–98, 2003.

[DSSSC08] John Duchi, Shai Shalev-Shwartz, Yoram Singer, and Tushar Chandra. Efﬁcient projections onto the l1 -ball for learning in high dimensions. In ICML, pages 272–279, 2008. [Dun73]

J.C. Dunn. A fuzzy relative of the isodata process and its use in detecting compact well-separated clusters. Cyber. and Sys., 3:32–57, 1973.

[ET96]

B. Efron and R. Tibshirani. Using specially designed exponential families for density estimation. In Proc. of the Annals of Statistics, pages 2431–2461, 1996.

[EUR05]

EURODIS. Rare diseases: Understanding this public health priority, 2005.

[FLG00]

G.W. Flake, S. Lawrence, and C.L Giles. Efﬁcient identiﬁcation of web communities. In KDD, pages 150–160, 2000.

[FM06]

Shai Fine and Yishay Mansour. Active sampling for multiple output identiﬁcation. In COLT, pages 620–634, 2006.

[Fun01]

Glenn Fung. A comprehensive overview of basic clustering algorithms, 2001.

[GGP07]

Gianluigi Greco, Antonella Guzzo, and Luigi Pontieri. An information-theoretic framework for high-order co-clustering of heterogeneous objects. In SEBD, pages 397–404, 2007.

[GKT05]

D. Gibson, R. Kumar, and A. Tomkins. Discovering large dense subgraphs in massive graphs. In VLDB, pages 721–732, 2005.

[GLZ+ 07]

B. Gao, T. Liu, X. Zheng, Q. Cheng, and W. Ma. Consistent bipartite graph co-partitioning for star-structured high-order heterogeneous data co-clustering. In KDD, pages 41–50, 2007.

[GV04]

Hongyu Guo and Herna L. Viktor. Learning from imbalanced data sets with boosting and data generation: the DataBoost-IM approach. SIGKDD Explorations, 6(1):30–39, 2004.

[HAR68]

P. HART. The condensed nearest neighbor rule. IEEE Trans. Inform. Th., 14:515–516, 1968.

[HB97]

Thomas Hofmann and Joachim M. Buhmann. Pairwise data clustering by deterministic annealing. IEEE Trans. Pattern

132

Bibliography Anal. Mach. Intell., 19:1–14, 1997.

[HC07]

Jingrui He and Jaime G. Carbonell. Nearest-neighbor-based active learning for rare category detection. In NIPS, 2007.

[HC08]

J. He and J. Carbonell. Rare class discovery based on active learning. In ISAIM, 2008.

[HCN05]

Xiaofei He, Deng Cai, and Partha Niyogi. Laplacian score for feature selection. In NIPS, 2005.

[HXD05]

Zengyou He, Xiaofei Xu, and Shengchun Deng. An optimization model for outlier detection in categorical data. CoRR, abs/cs/0503081, 2005.

[HYKL04]

Kaizhu Huang, Haiqin Yang, Irwin King, and Michael R. Lyu. Learning classiﬁers from imbalanced data based on biased minimax probability machine. In CVPR (2), pages 558–563, 2004.

[Joa99]

Thorsten Joachims. Transductive inference for text classiﬁcation using support vector machines. In ICML, pages 200–209, 1999.

[Joa05]

Thorsten Joachims. A support vector method for multivariate performance measures. In ICML, pages 377–384, 2005.

[Joh67]

Stephen Johnson. Hierarchical clustering schemes. Psychometrika, 32(3):241–254, September 1967.

[KC07]

Yong-Deok Kim and Seungjin Choi. A method of initialization for nonnegative matrix factorization. In ICASSP, pages II–537–II–540, 2007.

[KM97]

Miroslav Kubat and Stan Matwin. Addressing the curse of imbalanced training sets: One-sided selection. In ICML, pages 179–186, 1997.

[KNRT03]

R. Kumar, J. Novak, P. Raghavan, and A. Tomkins. On the bursty evolution of blogspace. In WWW, pages 568–576, 2003.

[KSM00]

YongSeog Kim, W. Nick Street, and Filippo Menczer. Feature selection in unsupervised learning via evolutionary search. In KDD, pages 365–369, 2000.

[LJF02]

Martin H. C. Law, Anil K. Jain, and M´ario A. T. Figueiredo. Feature selection in mixture-based clustering. In NIPS, pages 625–632, 2002.

[LS96]

R.B. Lehoucq and D.C. Sorensen. Deﬂation techniques for an

Bibliography

133 implicitly restarted Arnoldi iteration. SIAM J. Matrix Anal. Appl., 17(4):789–821, 1996.

[LZWY06] B. Long, Z. Zhang, X. Wu, and P.S. Yu. Spectral clustering for multi-type relational data. In ICML, pages 585–592, 2006. [MMP02]

Pabitra Mitra, C. A. Murthy, and Sankar K. Pal. Unsupervised feature selection using feature similarity. IEEE Trans. Pattern Anal. Mach. Intell., 24(3):301–312, 2002.

[MN98]

Andrew McCallum and Kamal Nigam. Employing EM in poolbased active learning for text classiﬁcation. In In Proceedings of the 15th International Conference on Machine Learning, pages 350–358. Morgan Kaufmann, 1998.

[MNT04]

Kaj Madsen, Hans Bruun Nielsen, and Ole Tingleff. Optimization with constraints, 2nd ed., 2004.

[Moo91]

A.W. Moore. An introductory tutorial on kd-trees. Technical report 209, University of Cambridge Computer Laboratory, 1991.

[NW99]

Jorge Nocedal and Stephen J. Wright. Numerical optimization. Springer, August 1999.

[PKGF03]

Spiros Papadimitriou, Hiroyuki Kitagawa, Phillip B. Gibbons, and Christos Faloutsos. LOCI: Fast outlier detection using the local correlation integral. In ICDE, pages 315–327, 2003.

[PM04]

Dan Pelleg and Andrew W. Moore. Active learning for anomaly and rare-category detection. In NIPS, 2004.

[RR95]

D.A. Reynolds and R.C. Rose. Robust text-independent speaker identiﬁcation using Gaussian mixture speaker models. IEEE Trans. Speech and Audio Processing, 3:72–83, 1995.

[RRS00]

Sridhar Ramaswamy, Rajeev Rastogi, and Kyuseok Shim. Efﬁcient algorithms for mining outliers from large data sets. In Weidong Chen, Jeffrey F. Naughton, and Philip A. Bernstein, editors, SIGMOD, pages 427–438. ACM, 2000.

[Sco92]

D.W. Scott. Multivariate density estimation: Theory, practice, and visualization. Wiley-Interscience, 1992.

[SCZ98]

Gholamhosein Sheikholeslami, Surojit Chatterjee, and Aidong Zhang. WaveCluster: A multi-resolution clustering approach for very large spatial databases. In VLDB, pages 428–439, 1998.

134

Bibliography

[Set10]

Burr Settles. Active learning literature survey. Computer Sciences Technical Report 1648, University of Wisconsin– Madison, 2010.

[SFPY07]

J. Sun, C. Faloutsos, S. Papadimitriou, and P.S. Yu. GraphScope: parameter-free mining of large time-evolving graphs. In KDD, pages 687–696, 2007.

[SG08a]

Ajit Paul Singh and Geoffrey J. Gordon. Relational learning via collective matrix factorization. In KDD, pages 650–658, 2008.

[SG08b]

Ajit Paul Singh and Geoffrey J. Gordon. A uniﬁed view of matrix factorization models. In ECML/PKDD (2), pages 358– 373, 2008.

[SKW06]

Yanmin Sun, Mohamed S. Kamel, and Yang Wang. Boosting for learning multiple classes with imbalanced class distribution. In ICDM, pages 592–602, 2006.

[SPST+ 01] Bernhard Sch¨olkopf, John C. Platt, John Shawe-Taylor, Alex J. Smola, and Robert C. Williamson. Estimating the support of a high-dimensional distribution. Neural Computation, 13(7):1443–1471, 2001. [TKK01]

Simon Tong, Daphne Koller, and Pack Kaelbling. Support vector machine active learning with applications to text classiﬁcation. In Journal of Machine Learning Research, pages 999– 1006, 2001.

[Tom76]

I. Tomek. Two modiﬁcations of CNN. IEEE Trans. Systems, Man and Cybernetics, 6:769–772, 1976.

[TZCK09]

Yuchun Tang, Yan-Qing Zhang, Nitesh V. Chawla, and Sven Krasser. SVMs modeling for highly imbalanced classiﬁcation. Trans. Sys. Man Cyber. Part B, 39(1):281–288, 2009.

[VW09]

Pavan Vatturi and Weng-Keen Wong. Category detection using hierarchical mean shift. In KDD, pages 847–856, 2009.

[Was05]

L. Wasserman. All of nonparametric statistics. Springer, 2005.

[WC03]

Gang Wu and Edward Y. Chang. Adaptive feature-space conformal transformation for imbalanced-data learning. In ICML, pages 816–823, 2003.

[WJL+ 07]

Zhe Wang, William K. Josephson, Qin Lv, Moses Charikar, and Kai Li. Filtering image spam with near-duplicate detec-

135

Bibliography tion. In CEAS, 2007.

[WXWC07] Junjie Wu, Hui Xiong, Peng Wu, and Jian Chen. Local decomposition for rare class analysis. In KDD, pages 814–823, 2007. [YSZ02]

Dantong Yu, Gholamhosein Sheikholeslami, and Aidong Zhang. FindOut: Finding outliers in very large datasets. Knowl. Inf. Syst., 4(4):387–412, 2002.

[YZ05]

Raphael Yuster and Uri Zwick. Fast sparse matrix multiplication. ACM Trans. Algorithms, 1(1):2–13, 2005.

[ZBL+ 03]

D. Zhou, O. Bousquet, T.N. Lal, J. Weston, and B. Sch¨olkopf. Learning with local and global consistency. In NIPS, 2003.

[ZWG+ 03] Dengyong Zhou, Jason Weston, Arthur Gretton, Olivier Bousquet, and Bernhard Sch¨olkopf. Ranking on data manifolds. In NIPS, 2003.